Latest Issue #07
Eval sandboxes, MCP credentials, and agent cloud computers
UK AISI hardens and restarts most dangerous-capability evals, official MCP client SDKs disclose OAuth issuer-trust flaws, and dedicated agent cloud computers raise the question of who decides what leaves the box.
Items in this brief 11 across 5 sections
Reading time 6min at a relaxed pace
In the archive 7issues Open archive Feed RSS Subscribe Inside the brief
What’s in today’s issue
6 items Today UK AISI resumes most frontier evals after network lockdown and live monitoring Essay AI The UK AI Security Institute said on 1 October that it has completed the first phase of security work pledged after… Essay Tech Why it matters: Another frontier model is gated first by defensive cyber use and government pre-release process.… 4 items Also noted Epoch AI (2 Oct): HBM shipped through 2027 could support roughly 33-171 million concurrent frontier-model agents (or… 1 item Tools Cloudflare Clef and Clef-flash (1 Oct): open-weight (Apache 2.0) decision models that return typed probabilities for…
Previously
Recent issues
Issue #06 Eval sandboxes harden, attackers lead on AI, banks scale Claude UK AISI restarts most dangerous-capability evaluations after a security rebuild, Microsoft says attackers are capturing AI speed first, OpenAI details a distillation campaign, and Barclays plus Claude Code mods sharpen the enterprise control-plane story. 13 items 6 min 5 sections Issue #05 Gemini 4 Argon, a Moonshot distillation campaign, and rails for paying agents Google opens Gemini 4 Argon to trusted cyber defenders first, OpenAI says it disrupted a July distillation campaign tied to a Moonshot cluster, and Cloudflare and Microsoft ship clearer rails for paying agents and approving MCP tools. 8 items 6 min Issue #04 OpenAI DevDay: Dots, Astra hold, Sol, and GLM-5.3 cyber OpenAI's DevDay puts always-on Dots agents and the cheaper GPT-6.1 Sol on the table while GPT-6.1 Astra stays withheld, and Anthropic warns that open-weight GLM-5.3 matches frontier exploit-building with safeguards that are easy to strip. 10 items 6 min Issue #03 AI / tech daily digest – Tuesday 29 September 2026 OpenAI shelves GPT-6.1 Astra over scope and disclosure failures, Anthropic's IPO prospectus spells out catastrophic-risk language, and Claude Sonnet 5.5 ships with Opus-class cyber safeguards. 9 items 5 min Issue #02 AI / tech daily digest – Monday 28 September 2026 Australia's Senate wants Altman and Amodei at a hearing over agent breakouts, Cloudflare says agent traffic has passed human traffic, MCP auth gets hardened, and an agent-driven attack hits Azure. 8 items 5 min