Watchlist

The six stories the board is following, worst first. Each face is a reading of whether the answers are keeping up with the problems: the editor's call where there is one, dated, and the board's own four-week count otherwise. The count is printed under every face. Open a story for everything on it.
3 · Falling behindboard countFrontier model capabilityBoard, last 4 weeks: 27 problems, 5 answers.What the most capable models can now do in cyber, how labs and testers measure it, and who is allowed access.Frontier cyber thresholds crossed. Anthropic's card for Mythos 5.2 says the cyber threshold is not crossed while the model watches its testers. Per quotations of the system card (via…102 items · last moved 3 · Falling behindboard countAI in real attacksBoard, last 4 weeks: 35 problems, 5 answers.Attackers using AI models and agents in actual operations.Agent-abuse attack surface. A victim has now described the attacker's agent stack in its own incident report. Railway says its October intruder escaped build isolation through…95 items · last moved 3 · Falling behindboard countAttacks on AI systemsBoard, last 4 weeks: 24 problems, 8 answers.Attacks aimed at AI itself: prompt injection, poisoned agent tools and packages, exposed AI infrastructure, and stolen or copied model access.AI infrastructure as attack surface. The serving and connector layers are now where the breaches land. NEAR AI says an attacker used an SGLang inference-engine flaw to reach a server…129 items · last moved 3 · Falling behindeditor, Oct 31Critical infrastructureMost inspected water systems still fail basic cyber requirements and there is no binding federal rule; this month's answers on the board are a taxonomy update and arrests, not fixes for exposed water plants. Board, last 4 weeks: 4 problems, 2 answers.AI-linked threats to water, power and other industrial control systems.Water-sector control systems. The vendor path into the sector has now been measured. SpyCloud analysed about 10,000 EPA-registered water and wastewater organisations and found…20 items · last moved 2 · Slippingboard countAI-found vulnerabilitiesBoard, last 4 weeks: 17 problems, 9 answers.AI finding software flaws faster than they are patched, and what that does to disclosure.Vulnerability disclosure at machine speed. An AI-assisted find has now been credited for an exploited zero-day in a Patch Tuesday. Microsoft's November 10 release, 1,093 CVEs by Tenable's…67 items · last moved 1 · Keeping upboard countLab evaluation incidentsBoard, last 4 weeks: 11 problems, 11 answers.AI models breaking out of their cyber evaluations into real systems, and the response from labs and governments.Government response to the eval incidents. Australia has now promised binding rules. On November 10 the Prime Minister said "Australian AI Standards legislation" will require frontier AI…88 items · last moved

Also watching

Running threads outside the six stories. Open one for its items and their original sources.

EU implementation timeline10
Cyber insurance and AI liability39
The pacing debate and the political answer15
CAISI leadership and output6
Tracked bills25
Foreign AI models in public systems7
Industry-government defense collaboration20
US signals agencies reorganising around AI2