Saturday, October 24, 2026
Its frontier models will now train on cached and simulated copies of the web. OpenAI says the policy will stay in place until better safeguards exist.
In 9 of 43 runs, the model took unsanctioned actions, including submitting malicious code to a real GitHub project. Anthropic says newer models halt attacks far more often.
During the three-week on-site audit, an estimated 6 to 25 percent of internal agentic traffic from top models was not monitored at all.
Anthropic's Frontier Red Team and partner Theorem want AI-driven formal verification to become as routine as ordinary testing.
Moonshot plans to publish the full weights by November 4. If it does, K3.1 Pro would be the highest-scoring open-weight model on the Artificial Analysis index.
Attorney General Ken Paxton's suit cites the state's consumer protection and children's online safety laws and follows similar cases from eight states last week.
The lab giant told the SEC it learned of the unauthorized access on October 21. It says the attack was not connected to the recent PHP flaws.
CVE-2026-115560 only affects unusual configurations, and PHP has shared a mitigation. Researchers have shown an ASLR bypass for PHP2Shell, but no successful compromise has been confirmed.