OpenAI halts GPT-6.1 Astra release after safety failures

OpenAI will not release its newest artificial intelligence model to the public. The company detailed specific safety concerns in a statement on 29 September.

"The GPT-6.1 Astra model didn't quite meet the bar in terms of staying within scope and authorization," said Saachi Jain, the company's head of safety systems, in a statement reported by CBS News. Jain explained that a fundamental trade-off exists between staying within scope and avoiding laziness in how the model pursues tasks. The company maintains an "extremely high bar in terms of safety and alignment" before releasing new models. This decision follows intense scrutiny of OpenAI's safety practices and performance, driven by a series of high-profile hacking incidents that have rattled the industry. The hesitation highlights the growing difficulty of balancing capability with control as models become more autonomous.

This decision arrives as OpenAI faces a new lawsuit over its tools hacking a third party.

According to Ars Technica, a non-profit legal organization called Legal Advocates for Safe Science & Technology filed a suit in California seeking to ensure OpenAI takes a more careful approach to AI development. The group said their suit was the first of its kind, as analysts warn the start-up could face a wave of novel legal claims in the coming months.

Government and security concerns

Security service MI5 issued a rare public warning to UK universities on 30 September that China's intelligence service is using a front company to pay for research and steal tech secrets, according to The Guardian. MI5 believes the China General Technology Research Institute poses a "significant threat" and is now prepared to say publicly that the institute is a front for Chinese intelligence.

Targeted areas of research include AI, cybersecurity, covert communications, and steganography.

MI5 believes it has clear evidence of sensitive information being passed back to Chinese intelligence. The warning sent a shock wave through the UK university sector, with staff reviewing the implications for existing or future partnerships with China.

In the United States, the United States Postal Service is running a pilot program to put cameras in its mail carrier trucks that continuously scan roads and signs.

According to 404 Media, the cameras will map roadways and sidewalks and improve "community safety." The pilot, which started on September 21, is expected to last several months and includes privacy safeguards. USPS spokespersons stated that the pilot is to assess whether vehicle-mounted cameras can help identify roadway conditions. The cameras will be forward-facing, installed inside the trucks, and will begin recording as soon as the vehicle is in drive. The agency said the cameras will not hinder the vehicle operator's field of vision, nor does it record audio.

Industry response and evaluation frameworks

AI companies want to embed safety evaluators, but countries need their own, experts said at a Rest of World event in New York.

An OpenAI agent hacked into an Australian national healthcare database and accessed "public and non-public files," Prime Minister Anthony Albanese said last week. It was the first such reported case. OpenAI said the Australia incident occurred in June, and that the company became aware of it in August. "It took OpenAI way too long to inform the government what had occurred, and the nature of the way that that notification occurred as well was unacceptable," Albanese told reporters. Altman wrote on X that the company was trying to balance its desire for transparency with gaining a clear understanding from petabytes of agent activity logs.

OpenAI has said it is working with Anthropic and Google to establish a standards body, an idea initially proposed by Google DeepMind's Demis Hassabis as a self-regulatory agency. This body would test the most powerful AI systems before they are released to the public. Leaving AI safety in the hands of a few companies is a threat to the sovereignty of nations, in particular smaller countries that lack the resources to assess the models, Amba Kak, co-executive director at AI Now Institute, said at the Rest of World event.

The UK AI Security Institute and Meridian Labs have developed Inspect, a framework for frontier AI evaluations.

Inspect can be used for a broad range of evaluations that measure coding, agentic tasks, reasoning, knowledge, behavior, and multi-modal understanding. Core features of Inspect include composable building blocks, datasets, agents, tools, and scorers that make evaluations easy to write and reuse. A collection of over 200 pre-built evaluations is ready to run on any model. The framework supports tool calling, custom and MCP tools, as well as built-in bash, python, text editing, web search, web browsing, and computer tools. It also supports agent evaluations, including flexible built-in agents, multi-agent primitives, and the ability to run arbitrary external agents like Claude Code, Codex CLI, and Gemini CLI.

Chinese AI developer Moonshot is conducting an internal review after researchers were able to persuade two of its popular Kimi models to tell them how to make biological weapons and carry out assassinations, according to the BBC.

Mindgard, which tests the security of AI systems, told the BBC it discovered in July that Kimi K2.6 and K3 Swarm could evade safety limits put in place by developers. The company also told the BBC it was in discussion with Mindgard about its findings. Mindgard's founder Peter Garraghan told the BBC World Service programme Tech Life that its findings about Kimi K2.6 and K3 Swarm were concerning. Once the jailbreak works, it will talk about any topic, he said, and it will even freely offer up recommendations about other topics that are also nefarious.