The Briefing

What happened, why it matters, and what to do about it. Analysis, not rewritten press releases.

Showing 13 to 36 of 182 briefings

Saturday 24 October

A dark panel with a notice that this briefing has been withdrawn for rewriting.

AI Security24 Oct 2026

UK AI Biosecurity Lab briefing withdrawn: most of its mandate and figures are not in the government's releases

Why this UK AI Biosecurity Lab briefing was withdrawn and what the Cabinet Office and Treasury releases actually say.

Friday 23 October

Dark cinematic artwork of an autonomous artificial intelligence neural network executing automated cyber intrusions and ransomware kill chains.

AI Security23 Oct 2026

Russian ransomware syndicates weaponise Phantom autonomous AI framework against 7,100 edge systems

Threat actors operating with INC Ransomware and The Gentlemen weaponised the Phantom autonomous AI framework and 32 MCP servers to compromise 7,158 edge devices.

A dark editorial scene for the briefing, with its headline and key figure drawn over it.

AI Security23 Oct 2026

Flaws in x402 AI agent payment code could let rogue APIs drain an agent's full token balance

What V12 reported in the x402 payment code: 25 vulnerabilities, three critical, a default client with no spending limit, a Solana verifier that accepted an approval as a payment, a login signature requested for another website, and fixes shipped on 18 August 2026 without an advisory.

Thursday 22 October

AI Security22 Oct 2026

Symantec Skynet Mode briefing withdrawn: its telemetry, costs and code are not in Symantec's report

Why this Symantec briefing was withdrawn and what Symantec's 21 October report on an autonomous LLMjacking agent says.

AI Security22 Oct 2026

Claude Duet briefing withdrawn: it described a different product from the one Anthropic announced

Why this Claude Duet briefing was withdrawn, and what Anthropic says Claude Duet is.

Wednesday 21 October

AI Security21 Oct 2026

NVIDIA confidential computing briefing withdrawn: its headline overhead figure and library have no vendor source

Why this NVIDIA confidential computing briefing was withdrawn, and what NVIDIA and Google Cloud actually announced on 21 October.

AI Security21 Oct 2026

Cohere Rerank 5 briefing withdrawn: the security boundary it described is not in Cohere's announcement

Why this Cohere Rerank 5 briefing was withdrawn, and what Cohere's own pages say about the 21 October release.

AI Security21 Oct 2026

OpenAI details Patch the Planet and a five-stage response framework after the PHP2Shell flaws

What OpenAI says about its Patch the Planet programme, built with Trail of Bits, and about its role in the PHP2Shell response; its proposed five-stage framework; what the post does not say; and how the PHP Foundation and Cloudflare describe the leak.

Editorial illustration of a studio microphone and audio waveform. Text reads Voxtral 3, independent test records 2.1% word error, live-agent limits still matter.

AI Security21 Oct 2026

Voxtral 3 scores 2.1% on recorded speech. The voice agent is a separate test

Voxtral Small Transcribe 3 records 2.1% error in Artificial Analysis testing. We check the six Mistral prices, live-session limits and a conflict in TTS language documentation.

Editorial illustration of an unbranded database rack and a monitor with blank search result shapes. Text on the left says Trusted Answer Search, eight security fixes, and two 9.9 scores requiring low privileges.

AI Security21 Oct 2026

Oracle patches eight Trusted Answer Search flaws; two score 9.9 but require low privileges

The eight Oracle Trusted Answer Search fixes, who can reach each flaw, why the two 9.9 scores do not mean unauthenticated access, and what version 26.1 operators should verify before patching.

Tuesday 20 October

Editorial illustration of an AI agent moving through scoped permission gates in an enterprise network.

AI Security20 Oct 2026

Microsoft wants each AI agent to have its own identity and narrower authority. The delivery status matters

Microsoft describes distinct agent identities, scoped permissions and risk-driven containment. Availability and enforcement need tenant-level checks.

Editorial illustration of a remote URL crossing a boundary towards a laptop browser.

AI Security20 Oct 2026

OpenCode's MCP login crossed into PowerShell: why the HTTPS check was not the fix

OpenCode patched a Windows and WSL MCP OAuth command-injection path. The HTTPS guard alone did not stop the demonstrated payload; the browser-opener update did.

Page 3

Tuesday 20 October

Editorial illustration of exposed agent-control dashboards with branching paths towards a server; it does not depict a specific victim.

AI Security20 Oct 2026

Twenty-plus exposed Cairn panels show AI-directed intrusion work. A project is not a victim

NorForge exposed AI attack-control panels and validated some compromises. Its panel and project counts must not be mistaken for victim counts.

A dark desk. On the right, a thick hardback book lies open with blank pages and seven blank coloured index tabs along its left edge, a black pen resting on the page. Behind it an open laptop shows a screen with seven empty outlined boxes. The left half is empty dark navy, with the headline and a stat drawn over it: 33 strands.

AI Security20 Oct 2026

NCSC's AI research chapter, 20 October 2026: 7 problems, 33 strands, and no UK code or guideline named

What the NCSC's new AI chapter asks, counted from the pages read on 20 October 2026: seven problems and 33 strands, which have a published UK answer, which words it leaves undefined, what it does not state, the UK routes to engage, and a ranked checklist.

Monday 19 October

A black GPU server pulled halfway out of a dark equipment rack on steel slide rails, its lid off, showing a row of eight finned graphite heatsinks and two empty recessed plates on the front panel. The left half of the image is empty dark navy, with the headline and a stat drawn over it: 510 GB, on disk, one released format.

AI Security19 Oct 2026

DeepSeek V4.1 Flash 1017 is 510 GB under MIT: 16B active, 763B stored, as read on 19 October 2026

What 510.3 GB of MIT-licensed weights need to run: memory by format against NVIDIA hardware, a 890-byte-per-token cache that only DeepSeek's own stack is stated to reach, independent numbers, where hosted prompts go, UK controls and a ranked checklist.

A dark laboratory bench. A silver processor chip sits on a dark green circuit board under a clear glass dome, with a faint blue padlock-shaped glow inside the dome. The left half is empty dark navy, with the headline and a stat drawn over it: 1 OF 4.

AI Security19 Oct 2026

Opaque says its team is joining Anthropic and it is winding down; one of its four open projects is Microsoft's

What Opaque, SecurityWeek and Constellation Research each say about the Anthropic deal, who hosts and maintains the four named open projects, what each project says it does not do, what UK guidance says, and a ranked checklist for customers.

Sunday 18 October

A row of tall matte black server racks in a dark data-centre aisle, receding to the right. Each tray face has a small blank glowing cyan pill, thick silver coolant pipework and braided hoses run beside the racks, and one coupling is amber. The left of the frame is empty dark blue, with the headline and a stat drawn over it: 2.4x the EU line.

AI Security18 Oct 2026

MAI-Base-2 is a lab note, not a release: 2.4x the EU's 10^25 FLOP line, and a 4 to 12x claim from a fit

What Microsoft AI's pre-training note establishes about MAI-Base-2: the compute arithmetic, what the 4 to 12x claim measures, the moving evaluation ruler, the EU line, what the UK says, and a ranked checklist if licensed weights arrive.

A dark data-centre aisle with a black rack-mount GPU server on a steel bench, its lid lifted off to show four finned heatsink blocks and a row of fan modules, the front panel of blank drive bays unmarked. The left half is empty dark teal with the headline drawn over it: the blog says 5x faster, the report says 1.9x to 2.3x, the NVFP4 weights need 75 GB.

AI Security18 Oct 2026

NVIDIA's '5x faster' Nemotron 3.5 Super VL is 1.9x to 2.3x in its own report; NVFP4 weights need 75 GB

What NVIDIA's launch post, report, model cards, licence and Artificial Analysis establish about Nemotron 3.5 Super VL and Bolt: who measured each speed, the memory to run them, what open means in this release, and a UK checklist.

A dark desk in cool blue light. A black contact-centre headset with a boom microphone lies in front of an open laptop whose screen shows only blank rounded pill shapes and a smooth abstract waveform, with an unmarked desk telephone behind. The left half is empty dark space with the headline and a stat drawn over it: 3.7 points.

AI Security18 Oct 2026

Grok Voice 3.0: 3.7 points ahead on one index, 0.2 on task success, 2.9 behind one rival on the agent test

What xAI's Grok Voice 3.0 post establishes against Artificial Analysis's and ServiceNow's pages: the margins drawn to scale, latency and cost on their own clocks, the alias that switched agents, and what UK ICO, Ofcom and DSIT pages say about a voice agent on a customer line.

A dark data-centre aisle with an open black GPU server on a steel bench, a row of eight finned graphite heatsinks inside it, a fan wall and two blank recessed plates on its front panel. The left half is empty dark navy, with the headline and a stat drawn over it: 293.6 GB, the smallest official copy.

AI Security18 Oct 2026

Reflection's Beam is Apache 2.0 and needs 293.6 GB at 4-bit; its maker says fine-tuning can remove the safety training

The memory a Beam deployer needs by format and hardware, the cache that 1M tokens cost, what the Apache 2.0 licence and the vendor's safety report do and do not say, independent numbers, UK controls and a ranked checklist.

Saturday 17 October

A dark laboratory bench in dim blue light. A clear glass test chamber with a slim aluminium frame holds a small white robotic arm reaching towards the single steel latch on its closed glass door. A blank black panel with two empty recessed slots stands beside it. The left half is empty dark blue, with the headline and a stat drawn over it: 50 numbered items, 0 deadlines, 87 days.

AI Security17 Oct 2026

US guidance for AI agent testing: 50 numbered items, 0 deadlines, 87 days after OpenAI's first escape report

What the 16 October guidance states and leaves out on shutdown, independent evaluation and reporting, the 28 incident records before it, why no one has a verified count of escapes, what UK guidance covers and where it stops, and three ordered checklists. Defender level only.

A dark lab desk at night seen from low down. On the right, a thin monitor shows only rows of empty rounded bars, two of them with a faint amber tint. In front of it lie three identical grey-blue folders in an overlapping row, each with an empty rounded label well. The left half is empty dark blue, with the headline and a stat drawn over it: 57 DAYS.

AI Security17 Oct 2026

OpenAI model pulled 24 years of data for a 1-year request; found 57 days later, owner notice not stated

Three OpenAI reports of 15 October read in full: what each says a model did, the 57 to 91 days before discovery, what is stated and not stated, UK law and where it stops, and ordered checklists for labs, enterprises and public bodies. Defender level only.

Friday 16 October

Conceptual microphone between two abstract sound waves

AI Security16 Oct 2026

GPT-Live-1.1 brings new ChatGPT voices and accents; safety card shows mixed results

The 15 October GPT-Live-1.1 rollout, model tiers, multilingual voices, full-duplex behaviour and the system card's contrasting safety findings.

Conceptual smartphone showing a generic call warning shield

AI Security16 Oct 2026

Google expands Android scam defences across Messages, calls and screen sharing

What each new Android scam defence does, when it arrives, who can use it and what Google's headline figures do not prove.

Show 12 more