Latest AI news — September 18, 2026

Hacktron: broke into OpenAI with Claude — ethical bug bounty. Hacktron AI’s three-person team used Anthropic’s Claude in OpenAI’s bug-bounty program: they chained two critical bugs, reached several OpenAI employees’ ChatGPT/Codex accounts, then proved access to the company’s internal GitHub code with a harmless pull request. Entry was through HEIF/HEIC image conversion on the Discourse-based community forum (outdated libheif). Researchers say Claude Opus 4.8 could not yet write a working exploit; within hours of Opus 5’s release, it could. OpenAI and Discourse patched; the bounty was $6,500. Source: TechCrunch, The Guardian.

Anthropic: quietly stood up a Bay Area wet lab — physical biology, not only in silico. Reuters reports Anthropic runs its own wet lab near San Francisco; Eric Kauderer-Abrams (life sciences) confirmed that the “ultimate test” is still the real lab for now. Goal: rare / “undruggable” diseases, Claude-directed robotic experiments with limited human intervention under human oversight. No clinical trials yet, so they do not compete with pharma; partners include Genentech, BMS, and Novo. Source: Reuters.

Anthropic: Claude “leads” 26% of AI R&D — new public metrics. Anthropic published its R&D Automation Index (Epoch AI AL scale): in August 2026 Claude is not fully autonomous (AL5) on any measured sub-area, but at AL4 (“leads”) it leads 26% of the work; “AI collaborates” or above is >90%. On the internal platform ~30 000 agents at once; of more than a billion decisions, about 1 in 47 000 was blocked. In one July week ~6% of AI R&D compute went to safety work (~12% within AI-driven R&D, conservative). They plan to embed third parties to check the metrics. Source: Anthropic, Reuters.

OpenAI: Astra for Law — GPT-6 Astra legal foundation layer. On September 17 OpenAI introduced Astra for Law: GPT-6 Astra + a legal search index + tools and controls tuned for legal analysis. API customers (e.g. Harvey, Legora) can build on it; 26 new ecosystem plugins (Relativity, Clio, and others). With Free Law Project / CourtListener material, the search index covers more than 230 million URLs of US case law / statute / regulation sources. On the Vals AI Legal Research Bench validation set (200 questions), at highest reasoning it scored 54.0% overall correctness versus 38.7% for web-search GPT-6 Astra. Source: OpenAI.

Amazon enters the AI safety debate — tests yes, industry slowdown no. An Amazon spokesperson told Reuters: models should ship when they are ready and safe — “rigorous testing and strong safeguards”; progress and safety are not a trade-off. Industry and government should work together on defenses. The company did not say it would join labs’ coordinated pacing line (Amodei / Altman / DeepMind / Microsoft). Source: Reuters.