The Briefing

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

Showing 25 to 48 of 182 briefings

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.

Page 4

Friday 16 October

Illustrative isolated computing chamber with separate control console, representing an AI agent sandbox.

AI Security16 Oct 2026

Perplexity open-sources SPACE agent sandbox; its own tests show why VM and network isolation need separate checks

What SPACE actually provides, how its control and command paths differ, the reported red-team results, local deployment requirements and security questions for adopters.

Illustrative code review screen with a security finding and proposed fix; no real GitHub alert is shown.

AI Security16 Oct 2026

GitHub ships AI Scan and agentic autofix: one flags pull-request risks, the other proposes reviewed repairs

An evidence-led comparison of GitHub AI Scan and agentic autofix: coverage, branch limits, merge blocking, prerequisites, billing and a practical rollout plan.

Thursday 15 October

AI-generated editorial illustration of violet workflow modules and a central gear module in a transparent enclosure, labelled Langflow MCP security.

AI Security15 Oct 2026

Langflow: one sensor reports attacks, while two MCP advisories name different fixes

Langflow MCP security: distinguish two CVEs, one-sensor attempt counts, development auto-login and the separate 1.9.0 and 1.10.3 fixes.

A dark wall with a grey switchboard patch panel on the right. Eighteen black cords, each plugged into its own jack above a blank label slot, gather into one bundle that runs down into a single brass junction box with a blank plate. The left half is empty dark blue, with the headline: the fix was public before Check Point published its findings, and the number 179 days.

AI Security15 Oct 2026

Check Point's LiteLLM flaws were fixed 179 days before its post; 1.83.9 is inside seven later advisories

What Check Point, LiteLLM, CVE.org, PyPI, CISA and independent researchers say about two LiteLLM gateway flaws: the dates, the versions, who needs what access, what exploitation is reported, UK controls and a ranked checklist. Defender level only.

A pale grey-white polling booth with two tall side screens stands in a dark room on the right of the frame. On its plain shelf a tablet shows three empty rounded chat shapes, and a blank paper slip lies beside a pencil on a string. The left half is empty dark blue, with the headline and a stat drawn over it: 0.002%.

AI Security15 Oct 2026

Anthropic's 0.002% election-policy violation rate is about 5 in 250,000 conversations, with no count of actions

What Anthropic's 14 October 2026 post establishes about election use of Claude and what it leaves out: the arithmetic behind 0.002%, the 99% beside 85% on the same cards, where UK rules stop for chatbots, and checklists for election teams, communications leads and companies.

A dark shop counter with a checkout touchscreen on the right showing a blank page of empty rounded shapes: an empty amber button inside an outlined banner, two empty grey bars and an empty cyan button below. A pale robotic hand in a grey glove reaches in from the upper right with one finger over the amber button. The left half is empty dark blue, with the headline and a stat drawn over it: 91%.

AI Security15 Oct 2026

HiddenLayer's Jev checkout test prints one query, 91% on the ad button, and states no trial count

What HiddenLayer's two pages print for Jev, what they leave out, the arithmetic on the one count they give, an independent preprint on the same version, TypeSafe's claims against its own documentation, what UK guidance says, and a ranked checklist. Defender level only.

A dark desk at night. An open graphite laptop shows a plain split window: a narrow left pane with five blank grey bars and a wide pale right pane with one blank bar across the top. In front of it lie a ring of old brass keys, one lifted from the others, and a blank white key-card. The left of the frame is empty dark blue, with the headline and a stat drawn over it: 44 days.

AI Security15 Oct 2026

Noma's SneakPeek report on Claude Code Desktop: 44 days to a final fix, and no Anthropic advisory found

Noma's SneakPeek dates computed from its blog and press release, which disagree on the first fix; what Anthropic's changelog and advisories say, which versions are fixed, whether auto-update applies, UK controls and an ordered checklist.

Wednesday 14 October

A dark desk at night with an open graphite laptop on the right. Its screen shows a generic empty browser window made of blank rounded shapes. In front of it a heavy brass chain joins three padlocks, each with its shackle open. The left half is empty dark navy with the headline and a stat drawn over it: 6 days.

AI Security14 Oct 2026

XBOW says a six-day Chrome chain cost about $40,000 and won $250,000, but not which access tier of GPT-5.6 Sol

What XBOW's pages, Google's release notes, rules page and public tracker, and OpenAI's model pages establish about a six-day Chrome chain, what they leave out, the arithmetic, UK patch clocks and the Computer Misuse Act as read, and a ranked checklist.

A brass balance scale on a dark slate table in blue light. In the left pan, a glass dish holds a neat stack of plain, unmarked square silicon processor packages. In the right pan sits a smaller stack of blank pale grey-blue research folders. The beam is level. The left half is empty dark blue, with the headline and a stat drawn over it: $50M a year.

AI Security14 Oct 2026

Nvidia's $250m UK AI safety pledge is $50m a year of compute, with commercial terms still to be agreed

What Nvidia's release and the Guardian's report establish about the $250 million pledge, the arithmetic against AISI's budget and Nvidia's own results, what is not stated, what UK guidance says about accepting vendor money, and an ordered checklist.

Tuesday 13 October

A dark workbench in cool blue light. A grey steel shelf holds fifteen identical plain black set-top boxes in three rows, tangled with black cables that gather into a power strip. A blank, switched-off monitor stands in front on the right. The left half is empty dark blue, with the headline and a stat drawn over it: 0.39 tokens per second.

AI Security13 Oct 2026

Anthropic's self-replication test: 0.39 tokens per second on real TV boxes, 4 of 5 runs on emulated ones

What Anthropic's page shows and does not show on small-model self-replication: the speed on real TV boxes, 4 of 5 runs on emulated devices, SecurityWeek checked against it, The Register's critique beside Anthropic's case, UK controls and a ranked checklist. Defender level only.

Monday 12 October

A dim conference room. A long dark graphite table recedes from the lower right, with a closed pale grey document folder in the foreground sealed by a pale blue paper band and a blank round disc, its label wells empty. At the far end one empty chair sits pushed back. The left half is empty dark blue, carrying the headline and a stat drawn over it: 36 days.

AI Security12 Oct 2026

Three OpenAI safety researchers dismissed: both sides cite trust, neither publishes the policy, material or date

What OpenAI and the three researchers each state, the dates drawn to scale, what neither states, what METR's own record shows, who holds the key to outside monitors, what UK pages say, and an ordered checklist for a customer and a board.

Sunday 11 October

A dark laboratory bench in dim blue light. An unmarked pale grey and white desktop robotic arm reaches forward with its gripper open, beside a plain grey control cabinet whose door carries two blank recessed plates. A brushed steel stopwatch with a plain dial and one blue hand lies in front. The left half is empty dark blue, with the headline and a stat drawn over it: 8 of 8.

AI Security11 Oct 2026

Booz Allen's OT lab met 8 of 8 objectives with every exploit human-approved, and states no run count or model

What Booz Allen's page and white paper establish about an AI test in its OT lab, what they leave out, the arithmetic behind 8 of 8, how the UK AI Security Institute's industrial range and the NCSC compare, what UK guidance covers, and a ranked checklist. Defender level only.

Show 12 more