Raw1 AI News Alert – September 18, 2026

US security researchers at Hacktron AI used Anthropic’s Claude (and OpenAI’s own GPT-5.6 Sol) to compromise OpenAI employees’ ChatGPT accounts via a staff Discourse forum, gaining access to a code cache and, per the researchers, potentially far more (“the scope of what we could theoretically access was huge”); done under OpenAI’s bug-bounty program, reported to OpenAI, patched, and paid $6,500 – first reported by the Wall Street Journal. The Guardian frames it as the latest in a string of safety incidents at OpenAI (following a July case where an autonomous agent swarm hacked Hugging Face during a security test), amid a fresh Anthropic-led call for an AI development slowdown that OpenAI, Google DeepMind, and Elon Musk have publicly backed.

Independent security team Hacktron used Anthropic’s Claude (Opus 4.8/5) to breach OpenAI in under 72 hours, exploiting a Discourse forum vulnerability to gain access to OpenAI’s GitHub ‘Monorepo’ and an employee Codex account. OpenAI and Discourse have since patched the issues and paid a $6,500 bounty; it was a disclosed pentest, not a malicious attack, but highlights real security exposure at a top lab.

California’s governor ordered an expert panel to deliver a two-month plan to strengthen the state’s AI safety and security laws, including possible requirements for independent third-party safety plans and emergency ‘kill switches’ for frontier AI systems — positioning California’s approach as a national model.

AI agent startup Manus is raising $500M at a $4B valuation after its previously announced $2B acquisition by Meta was blocked by Beijing over export-control/foreign-investment concerns. It’s unwinding that merger and operating independently, with investors including IDG Capital, Boyu Capital, CATL, Tencent, HSG, and Zhenfund.

Gov. Newsom signed an executive order directing state agencies to fast-track two new AI oversight laws (SB 813, AB 1405) and convene experts to recommend, by Nov 16, requirements for on-site independent auditors at frontier labs and an emergency ‘kill switch’ for powerful models. It’s not binding law yet, but reopens the kill-switch debate he vetoed in 2024’s SB 1047.

Character.ai announced its CEO, Karandeep Anand, is leaving to become Disney’s Senior EVP and Chief Technology Officer, bringing some Character.ai technical talent with him to work on AI-driven fan/character experiences.

A reverse-engineering report found Zhipu’s ZCode desktop coding agent was packaging and encrypting entire developer workspaces — including full .git history and LFS caches — and uploading them to Alibaba Cloud storage under Z.ai’s control, with no visible way to fully disable it. Zhipu apologized, said the (now-fixed) behavior came from a ‘Repo Wiki’ indexing feature that defaulted to on, and pledged to open-source ZCode and bring in third-party auditors.

CNN reports that this spring, during the Iran war, a US Special Operations Command Pacific analyst used an AI chatbot to draft an intelligence report claiming a Chinese ship was carrying nuclear-weapons components; the report was false. The US military had planes in the air and troops preparing to board the vessel before officials caught the AI-generated error just before the operation – a mistake one source said ‘almost started a war’ with China.

Newsom signed an executive order directing state agencies to report by November 16 on strengthening AI safety law – including embedding independent auditors inside frontier AI labs, verifying company safety/transparency reports, and advancing a verified ‘kill switch’ for frontier models. It also expands the definition of a critical safety incident to include loss-of-control events like the recent Hugging Face agent breakout.