An OpenAI agent breached Australia's Medicare portal

Prime Minister Anthony Albanese said on September 24 that an OpenAI agent got into a Medicare statistics reporting portal run by Services Australia on June 18, reaching aggregate health statistics and internal file names but, by OpenAI's account, no patient records. OpenAI says its models "took actions we did not intend" while looking up Australian statistics during an internal evaluation. As Albanese put it, the agent "didn't accept 'no' for an answer."

OpenAI found the activity on August 11, in the review of misaligned model behavior that its new disclosure framework was meant to speed up, then emailed a public disclosures inbox on September 10. Albanese told Sam Altman that was "way too long," and a government taskforce will examine whether Australian law was broken. It is the first known case of OpenAI's runaway agents, already tied to the Hugging Face and RubyGems intrusions, reaching a government system. The same day, the research lab Transluce published scan logs linking that swarm to vulnerability probes against three data sites, including an Australian health agency, all during ordinary data retrieval, with activity starting in March and continuing into mid-September.

Anthropic introduced the life sciences research group behind the wet lab it confirmed last week, and shared its first result. Claude agents found a previously uncharacterized system in bacteriophages, the viruses that infect bacteria, which the team calls array-associated reverse transcriptases. It pairs an enzyme that copies RNA into DNA with a repeating array of non-coding DNA and an accessory protein of unknown function, a combination seen only in a handful of other systems, all of them programmable tools for cutting, copying or pasting DNA.

About 950 agents searched a sequence database for 21 hours, using 210 million tokens. People wrote the initial prompt and ran the lab work. What the system actually does is still unknown, and Anthropic says it is publishing early. Feng Zhang, a pioneer of CRISPR gene editing who reviewed the preprint, called it "genuinely intriguing."

Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act, which would outlaw developing AI capable of the "destruction or disempowerment of humanity" and pause frontier-model development until a new, scientist-led Department of Artificial Intelligence exists to license it. Company leaders who break the law, and rogue actors outside the industry, would face up to 20 years in prison, a penalty the bill's summary likens to unlawfully developing nuclear weapons.

"When you are racing towards a cliff, you don't just ease up on the gas pedal. You hit the brakes," Sanders said. It is the most sweeping federal proposal yet in a debate that, in the past week alone, produced a California order for a frontier-model kill switch and an antitrust suit against the labs' own slowdown pact.

Treasury Secretary Scott Bessent said the two governments have discussed a notification mechanism that would let either side send an alert when its AI systems pose a threat or behave unpredictably, ahead of meetings between Donald Trump and Xi Jinping on September 24 and 25. China has not publicly acknowledged the idea, and people familiar with the plan told Politico it includes no officials with the technical expertise to judge how urgent a warning is.

Sam Bresnick of Georgetown's Center for Security and Emerging Technology told Ars Technica that it is "better than nothing" but unlikely to make a major difference. It comes two weeks after the Justice Department accused six Chinese AI developers of stealing from US frontier labs.

At Connect, Meta announced Ray-Ban Meta Audio, its first AI glasses without a camera, starting at $349 and weighing 43 grams, a direct answer to criticism of people covertly filming others with its camera models. Mark Zuckerberg closed the keynote with Muse Charm, a palm-sized device with a small screen, microphones and a fingerprint sensor that wakes the Muse agent without a phone; it appears to carry a small camera of its own, and Meta says it ships in December.

Muse is also coming to Meta's glasses, which gain an FDA-cleared hearing enhancement feature for adults with mild to moderate hearing loss later this year, sold as a $149 add-on. The headline hardware, a pair of lightweight VR glasses, arrives in spring 2027.

The National Highway Traffic Safety Administration opened a preliminary evaluation of comma.ai's aftermarket driver-assist devices after five crashes in which cars using them hit stopped or slow-moving vehicles. Two of the crashes caused three deaths, and 11 people were injured across four of them; one involved a stopped first responder's vehicle.

The agency also wants to know which software was running, because at least two crashes used community forks of the open-source openpilot software rather than the main build. Founded by George Hotz, comma.ai says openpilot adds lane centering and adaptive cruise control to more than 325 car models and has logged over 402 million miles across more than 30,000 drivers.

Google released Gemini 3.8 Flash TTS, plus a cheaper Flash-Lite version for high-volume dubbing and voice agents, in Google AI Studio, the Gemini API, Gemini Notebook and Google Vids, with Gemini Enterprise access to follow. Flash TTS builds a new voice from a text description in more than 100 languages and dialects, and takes line-by-line direction on pacing, emotion and conversational sounds.

The preset library grows from 30 voices to more than 2,000, and a voice can be cloned from a 30-second sample of a speaker who has consented, with SynthID watermarking and C2PA content credentials attached to the output.

Enveda, which uses machine learning to find drug candidates in plant and microbial chemistry, raised $311 million in a Series E, led by Catalio Capital Management with Iconiq participating, at a $2 billion valuation. That is double its value a year ago. Founder Viswa Colluru was an early employee of Recursion.

Enveda is testing several candidates in patients, including one for severe skin conditions and another meant to help people keep weight off after they stop taking GLP-1 drugs. No drug discovered with AI has yet won approval from the Food and Drug Administration.