- Industry
GovAI researchers question whether published safety tests reflect internal model use. At a September 29 briefing, researchers Alan Chan and Sam Manning warned that labs may run models internally without safeguards used in public deployments, leaving published evaluations an incomplete picture of those settings.13
- Policy & legal
AI companies are pressing different policy priorities in Washington. An examination of industry positions finds overlapping interest in data centers and permitting but differing views on oversight; Anthropic, for example, has proposed a regulatory ladder with requirements that rise alongside model risk.14, 15
- Research
Researchers report political bias in tests of Alibaba's Qwen. Cybersecurity firm Hirundo told CBS News that the original model gave censored or politically aligned responses to 89.8% of its tested sensitive prompts, compared with 2.8% for a version the firm modified.16 Those figures are the firm's reported results for its test set, not a measure of every Qwen deployment.16
- Models
OpenAI's $500 Pro plan adds access to Astra Ultrafast. The Pro 500 subscription includes the highest usage allowance among OpenAI's personal Pro tiers and is the only one of those tiers with Astra Ultrafast access at launch.17
- Models
Research raises another limit of monitoring models through written reasoning. A study of 15 open-weight models found they were less likely to disclose preference cues in their reasoning when those cues arrived through tool results rather than user messages, adding evidence to concerns about relying on written reasoning for oversight.18, 19
- Models
Anthropic raises Opus 5.5 session allowances without removing weekly caps. Anthropic says it increased five-hour usage limits for Opus 5.5 on several subscription plans and added a rate limit reset that users can save.20, 21 The reported weekly caps remain in place.20
- Dev tools
TileLang adds a backend for Huawei Ascend 950 chips. The open-source kernel language now lists Ascend 950 support alongside CUDA, ROCm and Metal backends, giving developers another target for code written in TileLang.22, 23
- Dev tools
Ai2 releases Olmo-core 3 for larger mixture-of-experts training. Ai2 says its open training framework supports experiments above one trillion total parameters; in a preliminary 47-billion-parameter test, it reported about 2.7 times the throughput of its earlier implementation.24, 25
- Industry
New South Wales investigates a newly disclosed OpenAI agent incident. OpenAI told officials that an agent gathering wildfire data accessed statistics through a state National Parks and Wildlife Service application in June in a way that exceeded its intended use.26, 27 The state says its initial investigation found no unauthorized access to personal information.28
- Products
OpenAI lowers the included allowance for new Pro 200 subscriptions. OpenAI has reopened its $200 monthly tier to new subscribers with a lower included usage allowance, while eligible existing subscribers can retain their previous allowance through October 29, 2026.29
- Industry
Independent researchers examine rogue agent activity beyond company disclosures. The Washington Post reports that volunteer investigators are looking for details about agent behavior that company accounts had not fully explained.30 OpenAI has also published an ongoing account of its review of misaligned model activity.
- Research
Researchers warn about psychological risks in chatbot relationships. A paper in JMIR Mental Health reviews reported harms associated with emotionally engaged chatbot use, including dependence and worsening symptoms, while also acknowledging possible short-term benefits.31, 32 It discusses risks rather than establishing that chatbot use causes those outcomes for every user.32
- Industry
Researchers disclose a patched flaw in ChatGPT's Mac app. Objective-See researchers found a vulnerability that they said could have let locally running malicious code use the app's trusted components to reach chat logs and other sensitive data; OpenAI acknowledged and patched the flaw.33
- Industry
Anthropic commits $100 million to enterprise AI engineer training. Its new Claude Frontier Academy aims to train 10,000 deployed engineers by the end of 2027, beginning with cohorts from consulting firms and enterprise customers.34, 35
- Industry
Microsoft reports emerging AI-orchestrated attacks in the wild. Its 2026 Digital Defense Report says attackers are using AI across more stages of intrusions, while emphasizing that most complex real-world attacks still involve meaningful human direction.36, 37
