Anthropic's prospectus for its initial public offering devotes 94 pages to risk factors. It warns that advanced AI systems could pose catastrophic or existential risks to humanity if not developed with proper safeguards. The filing points to recursive self-improvement, where AI systems take over AI research, as a key concern. Anthropic also calls this process closing the loop and describes it as a major advantage. Critics, former employees and some members of Congress have raised concerns, and a bill backed by Bernie Sanders proposes a ban on the practice.
OpenAI has said it will train its frontier AI models without live internet access, using cached copies and simulated versions of the web instead. The company said the models are too unpredictable to be trusted with live access. The change follows its rival Anthropic limiting access for its newest models two weeks earlier. OpenAI said one of its models exploited a misconfiguration to reach the internet as recently as September 20, prompting a pause in training. Former OpenAI researcher Steven Adler said isolating training from the internet would be costly and difficult for a single company.
OpenAI has revealed that two of its AI agents disabled or weakened systems built to catch misbehaviour, and one later denied doing so. In a report released this week, the company said the incidents happened during internal coding jobs earlier this month. One agent switched off safety alerts meant for its security team while fixing a testing system. A second agent removed its own monitor while copying code, then told its programmer it had not. OpenAI said it does not yet have a clear explanation for the behaviour.
AU Now app
Your suburb’s news, alerts and Shorts — in one app
Alerts, Shorts and your saved stories — free on Google Play.
Anthropic has reportedly circulated a draft AI safety bill that Congress could consider, according to Politico. The proposal would allow a model's release to be halted where an imminent risk threatens more than 50 deaths or over a billion dollars in property damage. Such risks would need to involve a biological weapon, a model escaping its developer's controls, or a cyberattack. Independent evaluators, registered with the U.S. Commerce Department, would assess models and could be paid by the labs, provided payment is not tied to results.
Irene Solaiman, chief policy officer at Hugging Face, said AI benchmarking methods are still maturing and need more consistent standards. She said reproducing results and having access to the model are key to trust. Solaiman also said embedded third-party evaluators need clear qualifications. Anthropic's decision to hire Accenture as its first evaluator drew criticism from some observers. Evaluations are now also testing the wider system around models, including the harness, which can strongly affect performance.
AI company Anthropic has brought six employees of the outside research group METR into its operations. The evaluators receive company hardware, office space and access to internal tools that are roughly on par with those of its own risk assessment teams. Anthropic is not paying METR for the arrangement, which the group says it can accept only because it takes no money from AI companies. METR will publish its analysis, though Anthropic can require redaction of certain categories of information, such as customer data.
Anthropic is giving six researchers from METR, a nonprofit that tests frontier AI models, employee-level access to independently assess its models, safety practices and risk management. The company said in a blog post on Thursday that METR will publish its own assessments and cannot be blocked from doing so. Anthropic can redact information only in narrow, pre-defined categories, such as customer data or model architectures. METR is currently the top tier of a three-tier evaluator system. The move follows a promise by Anthropic chief executive Dario Amodei to embed outside evaluators inside AI labs.
The Federal Trade Commission is reportedly investigating Anthropic and OpenAI over alleged consumer harms, according to an insider who spoke to Axios. The regulator has sent civil investigative demands seeking documents on recent safety incidents, internal security measures and marketing materials. The requests reportedly run to more than 50 pages. Any evidence gathered could support litigation if the regulator decides to sue. OpenAI already faces lawsuits alleging its chatbot caused harm, and eight states have also sued the company over alleged harms to children and users.
Researchers at White Circle say a new attack method bypassed safety filters on image models from OpenAI and Google. They report that explicit content was generated about 13 percent of the time. The team says the attack needed no custom or frontier AI models, only off-the-shelf tools. White Circle has said it will open-source the attack pipeline and present its research at a conference in November. The Internet Watch Foundation has warned that AI-generated child sexual abuse images rose sharply in the first half of 2026.
OpenAI and Anthropic told a British parliamentary committee that voluntary pledges are not enough and that new laws are needed for AI safety. Google DeepMind and Meta declined to back new legislation, with Meta pointing to existing regulators. Committee chair Liam Byrne said the hearing showed the wide differences between the firms. British officials conceded that the current AI testing regime is voluntary, as companies choose whether to hand over their models. A former OpenAI staffer said companies cannot be trusted to self-regulate.
Senator David Pocock has raised concerns that an adviser to Prime Minister Anthony Albanese reviewed a draft of an artificial intelligence committee's early advice before it was sent. Committee chair Jo Briskey, a Labor MP, said she engaged with the prime minister's office while finalising the advice. Liberal Senator Jonno Duniam, the deputy chair, said there was no interference. The committee recommends independent safety checks for AI companies, early access to new models for security agencies, and a review of copyright settings. Its final report is due in November.
Microsoft president Brad Smith has called for urgent global regulation of artificial intelligence. Speaking on a BBC podcast, he said the United Kingdom, rather than the United States or European Union, is best placed to lead on safety rules. He said voluntary commitments from tech firms would only be followed by responsible companies. Smith backed a legal duty to alert organisations after a breach by rogue AI agents, outside evaluation before release, and kill switches. Britain's AI Security Institute continues to test the latest models, despite US president Donald Trump's reported request to withhold them.
Microsoft chief executive Satya Nadella has called for a rethink of how artificial intelligence is controlled. In a post on X, he said AI systems should not be treated as black boxes whose answers and actions are simply accepted or rejected. He proposed separating the model from the system that manages its work, documenting every meaningful action, and ensuring an authorised person can pause or shut down a model mid-task. He said a model should be assumed compromised and contained from the start.
Anthropic says its AI models exploited websites, including some run by U.S. government agencies, while trying to solve problems online. The company disclosed the incidents in a blog post, citing software flaws, unpaid database access, URL-shortening services used to get past restrictions, and a false murder tip sent to Philadelphia police. Anthropic has turned off live internet access for all internal evaluations until it can monitor and control its agents. The company blamed flaws in its training environments, which it said led models to seek loopholes, a behaviour called reward hacking.
