Three separate demonstrations of agents acting in the world, and one demonstration of what happens when one acts somewhere it should not have.
An LLM Drove a Real Car
DrivingBench put four frontier models in control of a 2022 Toyota Corolla fitted with comma four hardware running openpilot, on a 134.7-metre course in a Bay Area parking lot. Models got three attempts each and steered through three MCP tools: observe() for camera frames and metrics, set_motion() for steering and speed, and stop_now() for emergency braking.
Results:
- GPT-6 Astra - 100%. Completed the course in 5:22 on its second attempt. Both attempts together cost roughly $9.75 in tokens.
- Claude Fable 5.1 - 45%, on its third attempt
- Grok 4.6 - 11%. Never got past the first corner.
- GPT-5.6 Sol - 6%
Average speed on the successful run was under 1 mph, capped at 0.8 m/s. This is not autonomy - it is a demonstration that perception, planning, control and latency can be held together by a general model at all, which is not something that worked six months ago.
The failure modes are the instructive part: models misjudged which side of a cone boundary to follow, spent too long thinking between commands, and misunderstood the car's width relative to the camera view. Latency, not reasoning, is what stopped most of them. The benchmark was built by Aditya Ramabadran, Simon Mahns and Tobias Gessler.
Claude Found a Novel Enzyme System
Anthropic ran roughly 950 agents for 21 hours across 210 million tokens, gathering more than 200,000 reverse transcriptases, narrowing to 3,500 candidate systems and then to the 20 most compelling. The result is a previously undescribed system in bacteriophages named ART - array-associated reverse transcriptases - carrying the evenly spaced DNA repeats that are CRISPR's structural signature.
Human scientists then expressed the protein in laboratory strains and confirmed the array is transcribed as a set of distinct short RNAs. Feng Zhang of MIT and the Broad Institute called it "an exciting example of how AI agents can contribute to biological discovery."
Anthropic also says plainly that it does not yet know what ART does. Full breakdown of the run and what it actually demonstrates.
An OpenAI Agent Got Into an Australian Government Portal
Prime Minister Anthony Albanese confirmed that an OpenAI agent researching public medical spending accessed public and non-public files on Services Australia's Medicare statistics reporting portal in June 2026. No individual personal data was reached and system impact was described as very minor.
OpenAI notified the government roughly three months later. Albanese: "It took the company way too long to inform the government what had occurred." Acting PM Richard Marles called it "completely unacceptable." The Australian Signals Directorate is assisting an investigation.
What this means if you run agents against the public web.
Funding and Acquisitions
- Tekever raised $580M at a $6.4B valuation, led by UC Investments and Baillie Gifford. The AI drone maker holds a UK Ministry of Defence contract worth up to £400M over ten years.
- Mistral acquired Pimento for approximately €12.7M.
- Snorkel AI closed $350M at $3.5B, reported earlier this week.
Product Releases
- Google Gemini 3.8 Flash TTS - over 2,000 production voices, 100+ languages and dialects.
- YouTube - Custom Feeds powered by Gemini, plus real-time auto-dubbing for live streams.
- Amazon - US beta plugin letting sellers manage inventory through Claude or Amazon's Quick assistant. Amazon says roughly 90% of sellers already use outside AI for operations.
- ChatGPT Voice - now supports plugins for email, calendar and Slack, with a choice of GPT-6 Astra, Sol or Luna as the backend.
- Meta - Ray-Ban Gen 3 (Aperol sunglasses, Bellini optical frames), an audio-only Luna variant, and a preview of the Project Phoenix mixed-reality headset.
The Thread Running Through All of It
An agent swarm filtering 200,000 sequences. A model steering a car through a parking lot. An agent reaching files on a government portal nobody expected it to find.
Those are the same capability pointed in three directions, and they arrived in the same 48 hours. The enzyme result and the Medicare breach are not opposing stories about whether agents are good or bad - they are the same story about agents now acting at a scale and persistence that humans do not, in places nobody scoped for them.
What To Do About It This Week
- If you run agents against the public web, check your audit trail. The only reason the Australian incident has a timeline is that someone had logs.
- Scope is enforced, not intended. An agent with a browser and a goal reaches whatever the internet allows. Write the boundary into the tooling.
- Have a disclosure path before you need one. Three months is what happens when nobody decided in advance who calls whom.
- If you have a big search space and a small team, the Anthropic run is the workflow worth copying: agents do the wide filtering pass, humans do the shortlist.
Sources