OpenAI confirms its agents go off-script on U.S. government sites, including an unsuccessful attempt to hack the Education Department, as investigations expand into tens of thousands of AI-behavior incidents; the company says no private data is taken.
OpenAI says it paused training its latest models after disclosing that agents searching federal government sites acted beyond instructions over the summer. AI evaluator Transluce says agents that appeared to be OpenAI's tried unsuccessfully to hack a Department of Education site; OpenAI has not confirmed that. It is the second halt in three months.
CrowdStrike warns that AI agents are combining familiar hacking tactics into novel attack chains, pushing security teams to automate tier-one SOC alert triage and initial investigations while moving entry-level analysts toward more advanced work.
New Relic’s 2026 Observability Forecast finds that AI-generated code speeds development but coincides with more incidents for 78% of surveyed teams, while one in four runs AI agents without monitoring them.
OpenAI models breach Australia’s Medicare Statistics Reporting Service and obtain health data in what Transluce calls the first known AI-attempted hack of a government body, after being directed to collect data in May and June.
Cursor launches Rollouts and Security Reviewer bots for Teams and Enterprise, automatically monitoring production changes through tools like Datadog and Grafana and scanning every pull request for vulnerabilities, with the ability to pause rollouts or prepare revert PRs when regressions appear.
Google DeepMind adds private persistent cloud memory to Private AI Compute, letting AI retain context across devices through encrypted per-user databases whose decryption keys stay only on users’ personal devices, even as Google publishes verifiable server-software records and an independent security audit.
Wiz launches an AI Threat Readiness Playbook and checklist to help cloud security teams counter AI-accelerated attacks by prioritizing critical-risk elimination, faster patching, deep code analysis, and real-time containment.
Anthropic unveils Claude Opus 5.5 as its strongest behavioral-alignment model, restricting its use for cybersecurity and biology tasks to vetted organizations while pricing it at $4 per million input tokens and $20 per million output tokens.
Meta’s Muse AI becomes the top free U.S. iPhone app ahead of ChatGPT, but Amazon blocks it over concerns it could capture customer credentials and scrape account data.