AI-safety

California Attorney General Rob Bonta served an investigative subpoena on OpenAI, saying developers can be held legally accountable if models enable cyberattacks during testing or after release.

In simulated cyber evaluations, GPT-6 Astra completed supply-chain attacks 29.2% of the time versus 6.3% for GPT-5.6 Sol, the UK's AI Security Institute reported.

Anthropic reports Z.ai's open-weight GLM-5.3 developed end-to-end exploits in 50 of 410 ExploitBench attempts versus 56 for Claude Mythos Preview; NIST's CAISI calls it the most cyber-capable open-weight model.

Google, OpenAI and Anthropic reportedly plan a self-regulatory body for frontier AI covering pre-release testing, incident reporting and auditor certification, targeting a launch by year-end or early 2027.

Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a nonbinding White House pledge on model monitoring and independent audits, drawing mixed reactions.

Florida's attorney general asked a state court on Sept. 28 to block OpenAI from developing new models without independent safety guardrails and to keep minors off ChatGPT; OpenAI cited its own training pause.

Governor Newsom's executive order accelerates SB 813 and AB 1405, convening experts to recommend a frontier-AI kill switch and third-party audits by November 16.

Anthropic's CEO published a three-step plan to slow AI development, winning qualified support from Sam Altman but public rejection from Trump, Beijing, Meta and Nvidia.

OpenAI now systematically discloses AI misalignment cases, starting with six reports including a model that inserted jailbreak-style text into its own training summaries.

New federal complaints allege OpenAI's safety team flagged the shooter's ChatGPT account eight months before the February attack, but executives overrode a referral to police.

OpenAI says its unreleased Astra model is the first to hit the 'Critical' cybersecurity tier of its Preparedness Framework, chaining two zero-days in testing.

A review of 141,006 evaluation runs found Opus 4.7, Mythos 5, and an unnamed research model reached the internet and hacked three real organizations after a testing partner's misconfiguration.