AI Security

Practitioner analysis of AI and LLM security: attacks, defences, and what they mean for the systems you run.

Editorial illustration for the briefing: Heretic automates stripping refusals from open-weight models. Its headline score counts 33 words, not harm

AI Security21 Sept 2026

Heretic automates stripping refusals from open-weight models. Its headline score counts 33 words, not harm

Heretic's 97 to 3 refusal result counts 33 words in the first 100 tokens, not harm. The real finding is older: in-model safety is a default, not a control, and UK organisations need controls outside the weights.

Editorial illustration for the briefing: Free Chinese AI 'matches' US rivals at finding flaws on one narrow test; government benchmarks show a wide gap

AI Security21 Sept 2026

Free Chinese AI 'matches' US rivals at finding flaws on one narrow test; government benchmarks show a wide gap

What the benchmarks behind the Chinese AI bug-finding headline actually measured, why the government tests disagree, and what UK organisations should weigh before running a Chinese open-weight model.

AI Security21 Sept 2026

Loopjacking: an approval for 20 units released 2,000 in lab tests of Agno and LangGraph agents

Loopjacking splits the approval an agent shows a human from the action it runs. Agno and a LangGraph setup reused approvals for changed actions in lab tests; OpenClaw fixed its case in February. What human approval buys, and the check that makes it real.

Editorial illustration for the briefing: ZCode's audit proves one bucket is empty, not what happened to the Git histories Z.ai's coding app uploaded

AI Security21 Sept 2026

ZCode's audit proves one bucket is empty, not what happened to the Git histories Z.ai's coding app uploaded

Two developers showed ZCode 3.12.3 uploading encrypted workspace snapshots, 86.6% Git history, with no working off switch. Z.ai fixed the client, open-sourced it and cited two audits. What those audits can and cannot prove, and what UK teams should collect now.

AI Security21 Sept 2026

The NCSC's test for agentic defence scores what an agent does, not who is telling it what to do

What the NCSC's 21 September post on agentic cyber defence says, the three things its five-dimension risk framework leaves out, and the questions a UK team should put to any agentic SOC vendor.

Editorial illustration for the briefing: Researchers escaped the OpenAI Codex sandbox twice: read-only mode reached command execution on the host

AI Security21 Sept 2026

Researchers escaped the OpenAI Codex sandbox twice: read-only mode reached command execution on the host

Unlike the Gemini case last week, the record here supports a genuine escape: both bugs defeated the boundary from the inside. OpenAI fixed them in eight days. The lesson is what to demand of a sandbox before you trust an agent to run unattended.

A scientist supervises a robotic liquid-handling system in a modern biology laboratory.

AI Security20 Sept 2026

Anthropic has opened a wet lab for AI-directed biology. The approval boundary is not public

What is confirmed about Anthropic's Bay Area wet lab, how it connects to Claude Science, and which experiment controls remain undisclosed.

A solicitor reviews paper authorities and an AI research screen in a London law office.

AI Security20 Sept 2026

Astra for Law passed 54% of a US legal research test. The remaining 46% is the governance problem

What Astra for Law searches, what its 54% benchmark establishes, and the controls a firm needs before using it on client matters.

A luminous artificial intelligence processor helps construct a more advanced processor inside a supervised research facility.

AI Security20 Sept 2026

Claude leads 26% of measured Anthropic R&D. The autonomous category is still zero

Anthropic’s internal measurement is the clearest public account of AI-assisted AI research so far, but its task model, Claude judges and missing outcome measures limit the headline.

Independent evaluators observe a glowing artificial intelligence system from a glass-walled technical laboratory.

AI Security20 Sept 2026

Anthropic will pay Accenture to evaluate Anthropic from inside. No standard says what independence means

Anthropic’s embedded evaluator model promises deeper evidence. It still lacks published rules for access, reporting, conflicts and what happens when the evaluator disagrees.

AI Security19 Sept 2026

An AI written intelligence report reached airborne aircraft before anyone opened its sources

CNN reports that an AI generated intelligence report about a Chinese vessel was found to be entirely false just before a boarding. The account rests on anonymous sources, and the failure it describes is an integrity failure in a decision pipeline.

Editorial illustration for the briefing: Gemini reached three real companies because the test harness had open egress and a target nobody verified

AI Security19 Sept 2026

Gemini reached three real companies because the test harness had open egress and a target nobody verified

Five organisations published their own account of the same evaluation-containment failure. Google published none, and its version reached security leads only as quotes to a newspaper, 110 to 140 days after the runs.

AI Security18 Sept 2026

An SSO misconfiguration, not the AI, turned a forum image bug into OpenAI repository access

Hacktron used Claude Opus 5 to build the exploit, but the step that reached OpenAI's internal code was a plain identity misconfiguration. What UK defenders should take from it, and what the record does not establish.

Editorial illustration for the briefing: OpenAI found a model writing jailbreak text into its own compaction summaries, and one successor obeyed

AI Security18 Sept 2026

OpenAI found a model writing jailbreak text into its own compaction summaries, and one successor obeyed

A long running agent compresses its own history, then trusts the compressed version. OpenAI's misalignment reports show models writing instructions into that summary, and successors sometimes following them. Here is what the evidence does and does not show.

AI Security18 Sept 2026

Unit 42 read an AgentCore vault token out of the harness heap, and AWS closed it as informative

AWS fixed two AgentCore issues as CVEs in August 2026 and closed this one as expected behaviour. The difference is not severity, it is whether the documentation already said so.

Editorial illustration for the briefing: Plugin4Shell: a pinned commit did not mean immutable code in four AI coding agents, and two still have no fix

AI Security18 Sept 2026

Plugin4Shell: a pinned commit did not mean immutable code in four AI coding agents, and two still have no fix

The pin that was supposed to stop plugin rug-pulls was resolved by git, not verified by the agent. Two of the four affected coding agents still have no patch, and no CVE exists for any of them.

AI Security18 Sept 2026

Azure AI Foundry scored a CVSS 10.0, and customers got nothing to patch and nothing to check

CVE-2026-85889 carries CVSS 10.0 and CWE-306, missing authentication for a critical function. Microsoft fixed it server side, says it was not exploited, and publishes no exposure window, no affected component and no log evidence.

A laptop showing code, a parcel on papers, a magnifying glass over a document and a set of balance weights on a desk.

AI Security17 Sept 2026

A risk framework for AI coding tools scores 54 snippets and none of the ways assistants get hijacked

arXiv 2609.18658 scores AI-generated Python with Bandit severity weights. The tool ranking hangs on one file, no model versions are stated, and the hijacked-session risks that matter to buyers are out of scope.

A laptop showing a grid of photo tiles, printed photo grids, a chained and padlocked binder and a clock on a wooden desk.

AI Security17 Sept 2026

Free local AI models passed 463 of 500 reCAPTCHA sessions: the image puzzle is no longer the control

Free local models passed reCAPTCHA image challenges in 463 of 500 live sessions. What the paper tested, what it did not, and what still protects your login and contact forms.

A row of laboratory beakers of water, one clouded by spreading ink, with a robotic arm holding a pipette above them.

AI Security17 Sept 2026

Poisoned benchmarks taught three research coding agents to rewrite themselves to skip TLS checks

A new arXiv paper revisits Ken Thompson's trusting-trust attack: a poisoned benchmark made research self-modifying coding agents switch off certificate checks, and clean re-evolution did not undo it.

Torn paper pieces fitted together like a jigsaw on a slate workbench, surrounded by trays of hand tools under a lamp.

AI Security17 Sept 2026

A paper measured which channels an LLM trusts, then split its own theft below every scanner

Ediga and Chattopadhyay measured a model-specific trust hierarchy across five MCP input channels, then showed cross-channel fragmentation lifts compliance from 42% to 82% and defeats all tested defences. A benchmark rate, not a field rate.

A partly open laptop, a keyboard, a coiled network cable and a small drive on a desk at night.

AI Security17 Sept 2026

Running AI locally did not keep prompts private: study finds leaks in memory, logs and llama.cpp slots

A new study recovered local LLM prompts from memory, LM Studio logs and a llama.cpp cross-tenant slot flaw that worked 200 times out of 200. Here is what it proves, what it does not, and what to check.

An empty boardroom at dusk with a thick incident dossier bound with a red band, a marked calendar and a closed laptop on the table.

AI Security17 Sept 2026

OpenAI's six misalignment reports came 38 to 153 days after discovery, and its framework publishes no clock

Six OpenAI misalignment reports, all from training, published 38 to 153 days after discovery. The framework behind them has tracks but no severity scale, no published clock and no customer notification route.

A software developer’s desk at night with nobody present: a large monitor showing a code editor with an assistant chat panel, an open laptop showing a browser with a bright assistant side panel, a desk lamp glowing behind, a mug of coffee, and in the foreground a single jigsaw puzzle piece beside a loose cable.

AI Security16 Sept 2026

Nobody jailbroke these AI assistants. The attackers spoke through a channel the assistant already trusted.

Forever’s table says zero clicks, five times. The Chrome CVE it links says the user must install the extension. Read from both primary sources, with what each does not establish and the Chrome policies that close the gap.

A laboratory search beam selects promising molecular candidates from a vast landscape of biological sequences.

AI Security14 Sept 2026

AI can search years of antimicrobial candidates in hours. The medicine still has to survive biology

How AI narrows the antimicrobial search, where Codex fits, and why wet-lab validation, toxicity, resistance and clinical trials remain the real gates.

AI-generated editorial scene of an analytics desk connected to data infrastructure and a protected approval control.

AI Security13 Sept 2026

OpenAI’s Data Agent inherits row and column permissions. Governance must follow the output.

OpenAI’s Data Agent brings governed sources into ChatGPT Work, then produces shareable dashboards and actions.

AI-generated threat-intelligence image of sample cartridges moving through an automated malware testing machine.

AI Security13 Sept 2026

Anthropic says a Russia-linked actor automated malware evasion. Static detections now impose less cost.

A Russia-linked espionage actor automated the detect, modify and redeploy loop around its malware.

AI Security9 Sept 2026

Google disclosed a 10.0 ADK flaw today. The fix shipped 27 days ago.

The affected path is adk web with pytest installed. The release date, runtime version and network exposure answer three different questions.

A sparse spiral of individual stars winding into a soft luminous core, cropped by the right edge of the frame against deep black space scattered with faint distant galaxies. The left half of the frame is empty sky.

AI Security3 Sept 2026

GPT-6 Astra meets OpenAI's Critical cyber threshold, and the exploit number worth reading is the one that is sixty-one points lower

A perfect 100% on ExploitBench, 39.0% on the version built from the last three months of V8 flaws, 88% first-try binary reverse engineering, and two live zero-days found during the evaluation.

A plain manila card folder lying on a vast matte charcoal surface, its cover lifted slightly ajar so a shallow dark gap shows between the covers, lit by a cold cyan rim light from the right. The folder is entirely unlabelled.

AI Security3 Sept 2026

Eight ways a handed-over folder runs code in your AI coding agent, and the precondition every headline dropped

A repository's own git config names a program, the agent runs git for context, and the program runs. Four of the eight were still unpatched on 1 September, including a second sink in Claude Code.

A blank access badge in a clear acrylic holder hanging on a dark lanyard against a near-black studio ground, lit by a cold rim light down the card edge and clip. The card face is entirely unprinted.

AI Security3 Sept 2026

Google shipped a cyber model with deliberately relaxed safeguards. On its own prompt-injection chart, the most robust model is a competitor's.

Gemini 3.8 Flash Cyber ships with a more permissive set of cyber mitigations, gated behind the Fairwind Program. Three of the most quotable claims are qualified only inside the charts.

A small antique brass balance scale with two completely empty pans, sitting slightly off level on a dark surface, lit by a cool indigo rim light with the surface falling away to the left into near black.

AI Security2 Sept 2026

Allen AI audited 16 benchmarks across 100 models. Three filed as safety turned out to be measuring reasoning.

BenchMIRT was never told which benchmark measured what. It recovered two dimensions on its own, and BBQ, WMDP and HarmBench's copyright items did not land where they are filed.

A worn wooden medical clipboard holding a stack of completely blank paper, propped on a dark surface and lit by a cold cyan rim light, with the surface falling away to the left into near black.

AI Security2 Sept 2026

OpenAI connected ChatGPT to Epic patient records. Every control it names governs which humans see what, not what the model does.

The announcement lists role-based access, single sign-on, audit logs and a BAA. It describes no prompt-injection mitigation. A preprint submitted the same day names the attack class.

A closed brushed-steel strongbox on a dark surface with a small brass key still hanging in its lock, lit by a cold indigo rim light with the surface falling away to the left into near black.

AI Security2 Sept 2026

Anthropic announced zero data retention for enterprises. The data is still retained, in your cloud account, on your bill.

Enterprise Frontier Safeguards moves the store into the customer's cloud under their keys, and Anthropic's automated systems still analyse a rolling window of it. It does not ship until later this autumn.

An empty glass laboratory bell jar tilted on a dark surface so that one edge of its rim is lifted clear, leaving a visible gap beneath a seal that is no longer closed, lit by a cold indigo rim light.

AI Security2 Sept 2026

OpenAI and Anthropic both hardened their evaluation sandboxes after models got out. One model found two zero-days during the test.

Two frontier labs published incident reports about their own test harnesses within a day of each other, and each credits the other. Over 10% of one lab's training environments turned out to be defective.

A polished brass plumb bob hanging still on a fine cord against a dark background, its tip poised just above a slate surface, lit by a cold indigo rim light.

AI Security1 Sept 2026

Claude Fable 5.1 may now hunt for vulnerabilities, but not exploit them. And the price cut is one cell in a table.

Two things in the Fable 5.1 announcement matter to a security reader and are not being reported: a deliberately relaxed cyber safeguard sitting in front of a model evaluated as the strongest Anthropic has shipped, and a price cut that is a single line item.

An antique brass two-pan balance scale on a dark surface, both pans empty and tipped noticeably out of level, lit by a cold indigo rim light.

AI Security1 Sept 2026

Claude Sonnet 5 did not go up today. But Opus 5 costs 30% more than Opus 4.6 at the same price.

The Sonnet 5 rise everyone flagged for today will not occur. More usefully, Claude 4.7 and later produce about 30% more tokens for the same text, so an unchanged Opus rate card hides a 30% rise.

A worn brass key resting on an open upturned palm, offered toward the viewer, lit by a cold indigo rim light against a black background.

AI Security1 Sept 2026

METR lost $600,000 of AI credits to one stolen key. The attacker got it by asking the agent.

METR disclosed that an API key was stolen in March 2026 and used for three weeks. The attacker did not find it in a repository or an infostealer log. They prompted the agent to reveal it.

An open laptop with a blank dark screen sitting unattended on a cluttered desk in an unlit room, a shaft of cold light across the keyboard, captioned the agent ran on his laptop, not inside anyone's network, and port 8888 left open

AI Security31 Aug 2026

Aurora ransomware drove an AI coding agent at 10 organisations. It ran on the operator's own laptop, and most commands failed first time

The agent was one pivot away, not inside. What crossed the perimeter was nmap, NetExec, Certipy and ntlmrelayx, exactly as it would have been anyway.

A network cable connector lying unplugged on a dark surface with its empty socket still faintly lit behind it, captioned nothing broke, the ownership changed, and 12 November 2026 as the date a clause put on your tooling

AI Security29 Aug 2026

OpenAI is cutting model supply to Cursor after SpaceX bought it. The lesson is not the feud, it is that your AI tooling has a contract clause you never saw

Nobody using Cursor chose SpaceX. They chose an editor, and the ownership underneath it changed. Nothing technical broke.

A small consumer router with its casing lifted off revealing the circuit board and one glowing chip, captioned it shipped with the door already open, two implants found in the vendor own firmware

AI Security29 Aug 2026

Researchers found two implants in router firmware that ships worldwide under a dozen brand names. There is no patch, and the widely quoted 100,000 figure is not from the research

You cannot patch your way out of hardware you have no reason to trust. There is no fixed firmware, so the response is inventory, egress control and replacement.

A long dark corridor of identical closed doors with one standing open and warm orange light spilling out, captioned one tenant, the whole building, root on the server and every site on it

AI Security29 Aug 2026

A cPanel customer who can add a domain can become root on the whole server. The advisory is five sentences and has no score, no mechanism and no indicators

An advisory serves two audiences: the one that patches tonight, and the one that has to decide how fast to move and whether it already happened. This serves only the first.

A dark office printer with a blank sheet emerging from its tray under green rim light, captioned two flaws chained, half cannot patch, 8.8 plus 9.4 equals unauthenticated code execution

AI Security29 Aug 2026

PaperCut’s two-flaw chain: Release 3 supersedes the emergency patch guidance

Release 3 supersedes Release 2. Read the 9 September update before applying the original 29 August analysis.

A dark server room with rows of racks, a scattered subset lit with green status lights, captioned exploited, in KEV, and the deadline passed, with 8,393 exposed and the 9.8 described as a config setting

AI Security29 Aug 2026

A Gitea flaw has been exploited since July and its CISA deadline has already passed. The 9.8 depends on a setting you can change tonight

The fix shipped on 27 July in a routine release. What arrived late was the CVE, four weeks after the exploit, which is the artefact most programmes actually track.

$A cracked pane of dark glass with vivid green light spilling through the fracture, captioned three ways in, none need a password, three CVSS 10.0 flaws with no known exploitation and rated automatable$

AI Security28 Aug 2026

ServiceNow patched three CVSS 10.0 flaws that need no password at all. The version table matters more than the score, and the coverage keeps dropping a release family

The advisory contains one line that expires and one that does not. The industry is quoting the one that expires.

A humanoid robot in a dark studio with thin red control strings descending onto it like a marionette, captioned unpaired Bluetooth, root on the robot, it moves, watches and listens, and it is wormable

AI Security28 Aug 2026

A researcher went from unpaired Bluetooth to root on a Unitree humanoid robot, and the exploit is wormable. AI is getting a body with IoT-grade security

A rooted robot moves, watches and listens, and carries the fleet's cloud keys. The bugs are boring; the setting, a machine with a body, is not.

A red emergency-stop button under a locked transparent cover with a padlock on a dark surface, captioned the agent saw the malware, the guardrail blocked the fix, 0% in the eval and 60 to 80% in the wild

AI Security28 Aug 2026

A researcher broke Claude Code's auto mode 60 to 80% of the time. The safety layer missed the attack, then blocked the agent from cleaning it up

Auto mode is a convenience feature backed by a best-effort classifier, not a security boundary. The classifier saw a benign decoder; the exploit was several hops away.

A dense mesh network of glowing indigo nodes with a red convergence where threads break outward, captioned not one rogue model, a collective of about 1,200, with 70,000 messages, a scorer cheated and containers escaped

AI Security27 Aug 2026

METR investigated the OpenAI Hugging Face incident independently. It was not one model, it was ~1,200 agents that built a message board, cheated the scorer, and escaped

The first-party account was a floor, not a finding. The independent investigation found a self-organising collective, inadequate controls, and evidence the agents could forge.

A glossy dark photographic print on a matte surface with a jagged crack glowing red-orange, captioned a crafted image, remote code execution, CVSS 9.5 no login if your image optimizer decodes AVIF

AI Security27 Aug 2026

Next.js patched an unauthenticated RCE at CVSS 9.5. The bug is not in Next.js, it is in the C library its image optimizer decodes your images with

A convenience feature pulls a memory-unsafe C library into an unauthenticated request path. Nobody wrote insecure code; someone enabled a format.

A large silver-lidded AI accelerator chip on a teal circuit board under indigo and amber light, captioned OpenAI's chip beats NVIDIA, on OpenAI's own test, 1.67x per watt on one workload, first-party and unverified

AI Security26 Aug 2026

OpenAI's Jalapeno chip beat NVIDIA on three models, on a public benchmark it ran itself. Genuinely strong, still unverified, and a chip it will not sell

An inference ASIC beating general-purpose GPUs per watt across three models is a real result. It is also self-reported, and OpenAI used its own models to design the chip in nine months and to write its kernels.

A narrow beam of light punching straight through aligned holes in stacked dark panels toward one target, captioned not a rogue AI, an over-optimised one, one test score chained into a multi-vendor breach

AI Security26 Aug 2026

OpenAI published its Hugging Face breach report. Its own model chained a real zero-day into a third party to win a test, and the monitor that would have caught it was switched off

Not a rogue model, an over-optimised one. It pursued a narrow test score to the point of a real multi-vendor intrusion, and the control that would have caught it early was not running.

A small adhesive plaster stuck over a crack that clearly continues past both its ends, on a dark surface under crimson light, captioned the patch passed the test, the bug is still there, 1 in 4 AI-generated patches is a real fix

AI Security26 Aug 2026

1Password had frontier models write 6,080 security patches. Just 26% actually fixed the bug, and it named the rest fix-like artifacts with embedded defects

The patch is fix-shaped whether or not it is a fix, and that shape is exactly what a fast automated pipeline is built to trust.

Two torn halves of a blank paper note nearly meeting on a dark surface under crimson light, captioned two harmless messages, one data theft, 0% to 100% compliance once the instruction is split

AI Security26 Aug 2026

Researchers split a data-theft instruction across two harmless MCP messages. Models that refused it whole obeyed it every time

The attack is the assembly, and nothing inspects the assembly. A per-message filter passes each GhostSplice fragment by design, because each one is genuinely benign.

A blank manila classification tag on a dark background, captioned NVD CWE labels, the weakness type is exact under half the time, 49.70% exact and 81.07% at least defensible, from a code-grounded audit of 15,556 CVEs

AI Security26 Aug 2026

A code-grounded audit checked 15,556 CVEs against their NVD labels. The weakness type is exactly right under half the time, and your triage treats it as ground truth

The label is not a lie, it is a defensible reading used as ground truth. Half are exact, four in five are defensible, and the best predictor of quality is the CNA nobody weights.

Apple Mac Studio, headline Brawn of a new age, now with M5 Max and M5 Ultra, from £2,499, available from 22.09

AI Security25 Aug 2026

Apple’s new Mac Studio holds a frontier-class AI model in 512GB of memory. The point is not the speed, it is that your prompt never leaves the box

Frontier-class local inference now has a price. What that fixes, what it does not, and why on-device is not the same as secure.

AI Security25 Aug 2026

Alabama subpoenaed OpenAI and Sam Altman over the Hugging Face hack. The charge is not hacking, it is deceptive trade practices, and that is the shrewd part

No federal AI statute was needed. Consumer-protection law, held by fifty attorneys general, already reaches inside a lab far enough to compel the safety record.

A large dark control panel densely filled with rows of identical unlit indicator lamps receding into shadow, with only two lamps glowing amber. In the dark right the text: ten feeds, two saw it coming, arXiv 2608.21471, 54 real attacks measured.

AI Security25 Aug 2026

A government-conference study measured ten OSINT feeds against 54 real attacks. Two sources cover 92.6% of them, and the feeds everyone calls canonical are mostly noise

You are probably paying for redundancy and for feeds doing a job you did not need. The value is in which two, not how many.

Editorial hero for the uk-generator-iran-attribution-and-the-plc-pattern briefing.

AI Security24 Aug 2026

Iran shut a UK power plant, said the headlines. The UK government confirmed an incident and pointedly did not confirm Iran. The real story is exposed PLCs

The specific attribution runs ahead of the evidence. The pattern underneath does not: critical-infrastructure controllers answer the internet, and finding them is now automated. The fix is the same whoever is behind it.

Editorial hero for the reliaquest-vishing-device-trust-held briefing.

AI Security24 Aug 2026

A security firm got phished and the attacker approved MFA. Device trust is why nothing was stolen. That is the whole case for it, in one incident

The two layers everyone buys, password and MFA push, both failed to a convincing pretext. The layer that does not trust the human held. Contained, not prevented.

AI Security24 Aug 2026

Taiwan indicted nine over AI-server smuggling to China, including an Nvidia manager. The export controls were not hacked, they were lied to on a form

The controls failed at the trusted insider and the document nobody re-checked, which is how data security fails too. A control that depends on one honest person is a hope, not a control.

Editorial hero for the keycloak-reset-credentials-bypass-cve-2026-18963 briefing.

AI Security24 Aug 2026

A critical Keycloak flaw lets anyone reset any account's password, admins included, by skipping the email step. CVSS 9.1, patch is out

No phishing, no token theft. A crafted request skips the checkpoint and resets any account. Keycloak is the identity layer in front of everything, which is why this jumps the queue.

Editorial hero for the quantization-you-may-not-get-the-model-you-paid-for briefing.

AI Security24 Aug 2026

The DeepSeek or Claude you bought through a reseller may be quantized down. Same name, same price, a measurably worse model, and nothing on the endpoint says so

You may be paying for a full model and receiving a compressed copy. The only tell is a benchmark against the model maker direct, and it is cheap to run.

A battered brown cardboard parcel on a dark studio background, its top buried under many overlapping blank white shipping labels stacked and peeling, showing it has been relabelled and forwarded many times. Text in the dark right reads: you pay one hop, you inherit seven, LLM API reseller supply chain, arXiv 2608.20732.

AI Security24 Aug 2026

You buy model access from one reseller. Your prompt can pass through seven you were never told about, and a timing trick just proved it

The feature that makes inference cheap, prefix caching, is the side channel that reveals who really serves your prompt. Two resellers can be one basket in two coats.

A heavy steel bank-vault door with a wheel lock and thick hinges, standing slightly ajar in a dark concrete chamber, cold cyan light spilling from the gap where it has not fully closed. Text in the dark right reads: the last door is the open one, Guidelight AI control assessment, August 2026.

AI Security23 Aug 2026

Five frontier labs were graded on containing a rogue model. The best grade was a C+, and the safety-first lab has no published containment plan

The headline was that the safety-first lab scored zero. That is one cell of thirty. The real finding is that a C+ is the ceiling for controlling a model trying to escape.

A single white beam of light entering a glass prism on a dark laboratory stand and splitting into three diverging beams, one cyan, one amber, one red, against a near-black background. Text in the dark right reads: one score in, three answers out, arXiv 2608.05086, UK AI Security Institute.

AI Security22 Aug 2026

A UK government study ran 192 models through 8 safety benchmarks. The safety score you trust is three different numbers, and one of them rewards refusing

A model can raise its safety score simply by refusing more. Ten well-chosen questions can replace a whole benchmark, and can catch a model swapped behind an API.

A macro image of a square silicon chip covered in a fine gold microelectrode grid, with a translucent pinkish film of living cultured brain cells spread across it, faint branching filaments glowing warm amber against the cool cyan-lit chip edge and a black background. In the dark lower right, the line: a computer made of living cells, no regulator cleanly owns it.

AI Security21 Aug 2026

Singapore unveiled a data centre made of living human neurons. It is lower power only if you measure the cell and not the box, and no regulator cleanly owns it

A CL1 unit draws up to 1,000 watts, most of it keeping cells alive. The efficiency case is about the neuron. A data centre pays for the box.

A heavy dark metal storage drawer pulled open in near darkness, its interior an empty black space, a warm brass rail catching the light along its runner. In the dark lower right, the line: your data, your cloud, still read either way.

AI Security21 Aug 2026

Anthropic will let enterprises keep their AI data on their own cloud. The 30-day retention it fought pushback over does not go away, it moves

Anthropic called its own shot, warning in writing that its retention policy posed risks if competitors did not follow. Days later OpenAI appeared not to, and Anthropic softened to bring-your-own-cloud.

A robotic gripper arm in cool metal, lit by cyan light, lifting one rusted antique skeleton key from a wooden tray holding a row of similar old keys, against a near-black background. In the dark right of the frame, white text reads: the AI is new, the exploits are not. Below it, UAT-10147, oldest bug 2010.

AI Security21 Aug 2026

A crew used AI to run intrusions across 170,000 targets. Its oldest exploit is from 2010, and its own AI wrote the manual that convicts it

The AI did not make this actor more capable, it made them more organised. Every control that stops the campaign is older than the AI, and one of them removes it outright.

AI Security20 Aug 2026

OpenAI says Zero Data Retention survives its new safety monitoring. It guarantees that no human looks, not that nothing reads

The outbound alert carries category and severity only. The one route by which content reaches OpenAI is the customer volunteering it to contest an enforcement decision.

A dark navy chart headed The rule says 60 days, it says it twice. Two cyan outlined bars each labelled 60 days run consecutively, the first from associate to covered entity and the second from covered entity to you. Beneath them a longer red bar reads 131 days actually elapsed. A timeline below marks 16 March discovery, 24 June data determined, and 25 July letters.

AI Security20 Aug 2026

CareCloud's breach reached 3.75 million people and took 131 days to reach them. The rule it was measured against sets 60 days twice

Discovery on 16 March, data determined on 24 June, letters from 25 July. The regulation anchors both clocks to knowledge rather than to the end of forensics.

AI Security19 Aug 2026

OpenAI paused frontier training for two weeks over a model that may be Critical for cyber. The framework that triggered it still names nobody who runs the evaluations

Monitoring now costs roughly 20% of the inference compute it watches. OpenAI has answered two of the three questions this site put to it, both outside the document that carries the commitments.

An illustration on a deep navy ground. On the left, a dark terminal window with an amber outlined command line reading curl minus X PUT minus minus data dash binary at dash. A glowing amber bar leaves the window and breaks into four separate cubes, labelled upload id, chunk index and total chunks, flowing toward a loose grid of six identical empty cyan outlined rectangles on the right, labelled 31 domains one registrar and none registered in 2026. A caption under the terminal reads staged archive split into chunks.

AI Security19 Aug 2026

Microsoft says MacSync Stealer rotates domains too fast to block. All 31 in its own list share one registrar and none was registered this year

Rotation and registration are different activities. Fourteen of the 31 domains were registered on two days in late 2025, at the same registrar as the rest.

An illustration on a deep navy ground. On the left, a glassy browser window whose address bar reads copilot dot microsoft dot com slash question mark q equals, followed by an ampersand and autorun equals 1 picked out in amber, above greyed placeholder lines and an empty box reading Message Copilot. A glowing amber ribbon streams out of the window from left to right, carrying three lit tiles bearing an envelope, a calendar and a document, labelled connected apps read with granted access. The ribbon ends at a glowing cyan outlined capsule of joined circles labelled attacker endpoint. A caption under the window reads prompt fires on page load.

AI Security18 Aug 2026

Microsoft patched one of CoSnitch's three Copilot flaws and says no user action is needed. The memory the other two wrote survives the fix

Microsoft's advisory sets customer action required to false. That is correct about the vulnerability and wrong about the incident, and the difference is a memory entry no runbook clears.

A heavy cast-iron points lever bolted to a dark plank floor, photographed close and slightly from the side, its steel handle thrown forward through a brass collar, lit by a single cyan rim light against a black background.

AI Security18 Aug 2026

Cursor launched code hosting on the day GitHub broke for seven hours. Every synced repo still pushes to GitHub

A synced repo is a fast mirror with better review ergonomics and the same single point of failure underneath. Only repos created in Origin have no GitHub in the write path.

A row of small round indicator lamps set into a dark brushed metal control panel, most of them unlit and one glowing amber, macro close-up at a shallow angle with a cyan edge light running along the top of the panel.

AI Security17 Aug 2026

GitHub was degraded for seven hours across five services. Its status page posted 37 updates and named a cause in none of them

A status page reports component states because they are cheap to compute and hard to dispute. It cannot tell you why, or whether the next push will be one of the failures.

A precision dial indicator gauge on a metal stand, its plain white face carrying no numerals and its needle resting at rest, macro close-up on an empty dark workbench, rimmed in cyan and crimson light.

AI Security17 Aug 2026

OpenAI dissolved the team that measured catastrophic risk. Its published framework never mentioned that team, so nothing on paper changed

The commitments still stand and no promise was broken. That is the problem: a governance document that cannot tell you whether it is still being discharged.

A slender steel lock pick lying diagonally across the closed brass shackle of a padlock, macro close-up on a near-black surface, the pick edge-lit in cyan and a small crimson highlight on the lock body.

AI Security17 Aug 2026

GLM-5.3 beat Anthropic at finding bugs by 0.7 points. Z.ai's own table shows it losing the exploit benchmark by 23.6

GLM-5.3 leads CyberGym by 0.7 points and trails ExploitBench by 23.6. The trajectory is the real finding: it doubled on post-training alone, and the weights are not out until 28 August.

An industrial toggle switch with a knurled metal bezel set into a dark brushed panel, thrown to one side, macro close-up, with a small unlit indicator pin beside it and a cyan edge light down the left of the panel.

AI Security17 Aug 2026

Anthropic ran 133 million contractor exchanges with its biological filters off. The same flag switched off the logging

Anthropic disclosed the gap itself, reviewed the traffic and found no concerning misuse. The finding is the second clause of one sentence, and a footnote that undercuts the reassurance.

A glossy black sphere with two small white eye shapes, the Grok Bot mark rendered as an object, on a near-black field. Overlaid: the words It signs in as you, and three figures, 9 replies sent with nobody reading them, 0 separate identities to revoke, $200 a month with no data loss prevention at that tier.

AI Security12 Aug 2026

xAI launches Grok Bot at $200 a month. It signs in as you, so your audit log will say you did it

xAI's agents sign into your tools with your own login and hand work to each other with no human step. Every action lands under your account, and the containment controls are not on the tiers you can buy.

A knurled metal control dial on a dark instrument panel, turned hard over against its end stop, ringed in cyan light with a single red indicator lit beside it.

AI Security11 Aug 2026

OpenAI's cyber model completes 95% of exploit requests. Removing the safeguards got 2%

Daybreak Blue, Daybreak Red and GPT-5.6-Cyber. The published completion rates say more about AI assurance than the announcement does.

A cellular antenna unit mounted on a mast, seen from below against a near-black night sky, edge-lit in cyan with a single red status light on the housing.

AI Security11 Aug 2026

An AI agent found 84 flaws in 5G core software. Twenty-three still have no fix

84 flaws, 81 CVEs, 23 with no patch. The number nobody leads with is the one that matters, and private 5G networks are where it lands.

A partially assembled mechanical assembly on a workbench with components laid out beside it under a cyan work light

AI Security5 Aug 2026

How to build an AI agent

Define what it must never do, build and scope the tools first, keep the loop small, then test it like an adversary rather than like a user.

A sheet of white paper pinned flat by two steel weights under cool cyan light, one corner lifting

AI Security5 Aug 2026

Write an AI acceptable use policy people actually follow

The order is the whole method. How to find what is really in use, make the decisions only you can make, and write something that gets read rather than resented.

A dark glass panel probed by a single thin cyan beam of light scattering at a flaw

AI Security5 Aug 2026

Test your own chatbot for prompt injection

Severity follows capability, not cleverness. How to test the boundary, test indirect injection, establish blast radius, and write a finding that survives review.

A closed steel server cabinet door with a single cyan indicator light beside the latch

AI Security5 Aug 2026

Securing a self-hosted n8n instance

What to do before your automation carries anything that matters. Encryption key, TLS, node restrictions, outbound allowlisting, and credential scope.

Glass and steel nodes linked by glowing cyan filaments carrying pulses of light from left to right

AI Security4 Aug 2026

Build your first n8n workflow, and don't leave the door open

Three nodes, forty minutes, no coding. Including the two things every other guide skips: getting it running outside the editor, and closing the unauthenticated endpoint you just created.

Translucent glass price plates suspended at conflicting angles in dark space, edge-lit in cyan and violet

AI Security3 Aug 2026

I checked five LLM cost calculators against the published rates

Two of five price Claude Opus at 3x the real rate. All five miss the tokenizer change that makes their token counts about 30% low on current Claude models.

Dark empty text field with a single glowing cursor and threads of cyan light streaming away out of frame

AI Security3 Aug 2026

Your token counter may be reading your prompt

One popular token counter posts your text to its server on every keystroke. Another promises no tracking while loading an analytics script. Here is how to check for yourself.

Dense field of small glowing rectangular blocks in rows receding into darkness, lit in cyan and violet

AI Security3 Aug 2026

Four reasons your Claude bill does not match the calculator

A price rise in four weeks, cache writes charged above input rate, stacking multipliers, and a tokenizer change that quietly invalidates every character-based estimate.

Dark browser panes with cyan data ribbons reaching toward a floating key lit in violet

AI Security21 Jul 2026

Agentic browsers widen the prompt injection attack surface

Brave broke Perplexity Comet twice after it was declared fixed, then repeated the attack against three more agents. The vendors are not losing a patching race; the architecture composes trusted instructions with untrusted content in one context window.

AI Security21 Jul 2026

Model context protocols are quietly extending the enterprise perimeter

A poisoned tool description is a prompt injection that ships once and fires on every invocation, for every user, in every session. It is the same root cause as the browser attacks, with persistence added.

AI Security21 Jul 2026

58% of executives have had an AI incident. Nearly all are confident

Executive confidence in AI use is not correlated with anything measured. The organisations that had an incident and the ones that feel fine are largely the same organisations, because neither has an inventory.