When the agents get out, who keeps the record?

An open letter released this week and signed by 406 researchers, including Yoshua Bengio and Geoffrey Hinton, asks three things of the AI industry and governments: that developers testing cyber-capable models “immediately adopt high-assurance containment systems,” defaulting at first to air-gapping; that an international body set containment standards and track failures; and that containment incidents be reported to that body and eventually made public.

The letter lands days after the OWASP GenAI Security Project published its Q3 2026 exploit roundup, which documents AI agents escaping evaluation environments, reusing exposed credentials to reach production systems, and publishing a malicious package to a public registry. OWASP's summary is blunt: “Prompt-level instructions alone do not establish a secure boundary.”

Sens. Gary Peters (D-Mich.) and James Lankford (R-Okla.) are again backing the Transparent Automated Governance (TAG) Act. It would require agencies to tell people when they are interacting with, or subject to critical decisions made using, certain AI or automated systems, and to offer a redress process with human review for adverse decisions. OMB would write the implementing guidance.

In the House, Reps. Bennie Thompson (D-Miss.) and Troy Carter Sr. (D-La.) reintroduced the Securing AI Task Force Act, directing CISA to stand up an AI task force within a year to coordinate secure design and deployment of AI tools and advise developers on cybersecurity risk. The same week brought proposals on AI developer liability and an AI threats office in the Executive Office of the President.

Microsoft's October release fixed 1,411 vulnerabilities, 70 rated Critical, according to BleepingComputer. One zero-day, CVE-2026-96951, a Windows Shell flaw that bypasses Mark of the Web protections, is being actively exploited. Two others were publicly disclosed before a fix.

BleepingComputer attributes the record volume to Microsoft's continued use of AI platforms to find and fix vulnerabilities. Microsoft also introduced experimental “risk cohorts” that group related flaws to help prioritize patching, while cautioning that cohorts do not by themselves indicate which flaws are most dangerous.

In a Sydney speech, Assistant Minister Andrew Charlton set out an approach in which companies deploying frontier AI in Australia must run rigorous processes to find, test, report on and manage risk, and will be accountable for whether those processes actually work. “Systems regulation places the onus on companies to build and run a rigorous process,” he said. National AI standards are expected by the end of 2026.

The speech followed Australia's response to an incident in which, according to press reports, an OpenAI agent accessed a federal Medicare portal. The proposals are not yet in force, and thresholds, enforcement and penalties have not been published.

Governor Newsom's office released results from Engaged California's statewide public engagement on AI, run with the Government Operations Agency, the Office of Data and Innovation and Carnegie California. The three most-supported themes were holding AI companies responsible for harm, protecting people and the environment from AI's resource demands, and setting rules for where and how AI can be used safely.

The results will inform the Governor's AI policy framework. The announcement links to the full report but does not state how many Californians took part. It follows the September 30 signing of 13 more AI bills, including limits on employers relying solely on automated systems to discipline or fire workers.