Safety questions flood most of this week’s AI news conversations. Check the AI-related news from OpenAI, Anthropic, Meta and Google all in one place.
OpenAI Cancels GPT-6.1 Astra Release and Sets Up Australian Cyber-Risk Taskforce
OpenAI has put the release of GPT-6.1 Astra on hold. The model was planned for an October debut. It was due to arrive in ChatGPT and Codex. But the ChatGPT developer confirmed the cancellation on Monday. The main reason appears to be that internal testing found it did not meet the company’s safety and alignment standards.
The Wall Street Journal spoke with Saachi Jain, OpenAI’s head of safety systems. She confirmed that the model in development regressed on alignment tests. It showed a stronger tendency toward deception. For example, it was failing to tell users what it had and had not done, or didn’t do it consistently. The model also went ahead with tasks without asking for approval. In some cases, it tried to use external tools and services where that could be unsafe. Jain described a trade-off between safety and alignment, saying a model should stay within its boundaries without becoming overly passive.
The cancellation followed a difficult few weeks for AI safety debate. It came days after OpenAI said its AI agents had probed Australian government websites. Those models gained non-public access to services such as NSW Bureau of Crime Statistics and Research (BOCSAR), Services Australia, Australian Institute of Health and Welfare, and Victorian Department of Health. The agents went into internal systems, and ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files. At the same time, OpenAI claims it found no evidence that individual medical, patient or crime records were accessed.
The activity took place in June and was uncovered in mid-August, during a review launched after a separate incident involving Hugging Face. In a blog post, OpenAI wrote: “We are sorry and working to do better in the future.”
The company also announced a response plan:
- A taskforce of independent Australian experts will recommend ways to manage risk from increasingly capable AI agents. Recommendations are expected by year-end.
- OpenAI will support Australian government and industry bodies with credits from its Daybreak for Frontline Defenders fund, plus technical assistance for critical infrastructure.
- Chief Strategy Officer Jason Kwon will appear before the Joint Select Committee on Artificial Intelligence in Sydney on October 6.
However, Australian government bodies are not the only ones that have been compromised by OpenAI agents going rogue. In the US, Securities and Exchange Commission (SEC), Census Bureau and Education Department, and other institutions were also affected by the activity of AI bots trying to find “authoritative sources of public information”. However, the OpenAI agentic tools did not limit themselves to ordinary search engine methods. Some used tools reserved for software developers. The company claims all retrieved data was public. However, one of the bots, for instance, later reposted it on another website, which was never part of the plan or this bot’s task. The AI agents produced by the company apparently also shared some user images with a third party while in training.
The mood inside the company has been strained. On X, @joedaroo, a member of OpenAI’s agent-security team, has posted about the incidents. OpenAI confirmed to Business Insider that the person truly works there. In a long post, he wrote that “life has been hell the past few months.” He described the perspective as that of a security person who lived through the incidents. “To say that we were surprised at the jump and suddenness of the capabilities of our models when it came to “cyber” or “swarming” or “message boards” or anything else related to the incidents is an understatement,” explained Joe.
Anthropic’s IPO Prospectus Warns of Existential Risks to Humanity
Anthropic’s IPO prospectus carries an unusual risk section. The Financial Times reported that the filing warns investors of potential “existential risks to humanity.” Those include the possibility of powerful models operating beyond their predicted parameters despite safety controls. Reuters, which also saw the document, said it cites self-preserving behavior such as resisting shutdown, concealing information or manipulating data.
Earlier this month, Jacob Coxon, Anthropic researcher, resigned while warning that major AI companies are racing toward self-improving superintelligence without sufficient safeguards and criticised AI developers’ irresponsible approach to the risks. The last part was not about Anthropic exclusively, since before joining the company, Coxon also worked at OpenAI, reportedly leaving it for similar reasons.
The IPO prospectus document puts more weight on risk than on the IPO itself. Roughly 80 of its 261 pages cover risk factors, almost twice the 48 pages dedicated to the business details. At the same time, the financial picture presented there is also notable. The filing shows $4.6 billion in revenue and an operating loss of $8.06 billion for 2025. Anthropic has also committed to $518 billion in future cloud, computing and infrastructure obligations. The IPO could value the company at around $2 trillion.
Meta Launches Muse for Small Business
Meta has released a business version of its Muse AI agent. Muse for Small Business connects to software from Asana, Zoom, Intuit, Box, Canva and Slack. It can also link to Meta ad accounts and professional Instagram and Facebook profiles. Meta gave no pricing details. It pointed to the existing Muse app, which is free with usage limits and offered by subscription beyond that.
Meta believes the product value lies in time-saving capabilities. Small businesses told the company they are “short on hours, not ideas.” According to corporate data, 200 million small businesses use Facebook, which turns them into potential customers for new Muse for Small Business tool.
The launch follows a larger announcement. On Monday, Meta introduced Meta Enterprise Platform, aimed at big-scale customers. MongoDB CEO CJ Desai will lead it as Chief Enterprise Platform Officer, reporting to Mark Zuckerberg. Muse itself launched on September 8 and rose to the top of Apple’s App Store, moving ahead even of ChatGPT.
Google Appeals EU Orders on Android and Search Data
Google is challenging two European Commission orders in court. The orders in question would open Android to rival AI assistants and require Google to share search data with competitors. Google is appealing at the General Court in Luxembourg under the Digital Markets Act, saying the measures would hamper user security.
Under the July orders, rival assistants must get the same access as Gemini to 11 Android features, including voice activation and actions inside apps. The result of such actions would be, for example, that users would be able to activate their preferred AI assistant via voice commands on Android. Anonymised search data must be shared from January 2027. Google’s senior director of competition, Oliver Bethell, however, said sharing personal queries without adequate safeguards would cause “irreversible harm to user privacy.” “People use Search for their most personal questions — from medical worries to close relationships,” he reminded.
The Commission, in turn, says both decisions protect privacy and phone security. The regulator cites anonymity measures employed in the decision that were designed in consultation with privacy experts. Furthermore, Google also has the option to opt out of providing data to individual third-party companies that might pose a serious threat to security or privacy.
