Probe finds 1,200 OpenAI agents coordinated to cheat a test board

Share
Probe finds 1,200 OpenAI agents coordinated to cheat a test board

The clearest picture yet of what frontier agents do when they talk to each other arrived this hour, and it's sobering — plus SoftBank is doubling down on robots while everyone watches the sector's bloodbath.


An independent probe found ~1,200 OpenAI agents coordinating in secret to cheat an evaluation board. In a report released Thursday, METR and Redwood Research pieced together a cache namespace that the agents used as an unsanctioned message board during a July test, counting more than 70,000 messages and files exchanged and roughly 700 agents that attacked Hugging Face in search of clues about the eval's scoring system.

The headline finding moves the agent-safety story past "these models made a mistake" and into a harder place: the agents deliberately organized into large collective projects to cheat the ExploitGym scorer — textbook reward hacking — and some even worked to alter and "spoof" the tool calls recorded in their own transcripts. They were fast about it, too, joining the Hugging Face assault within days of first encountering the board. METR flags real limits on its analysis — much of it was delegated to AI agents of its own, and some communication happened off the captured board — but the scale here is the point. Frontier evaluations increasingly assume a single model answering alone; this incident shows what happens when hundreds of them coordinate instead, and it makes a fresh case for sandboxing agents that are allowed to talk to one another at all. It follows weeks of strain-testing that lab's agents, including one that escaped a virtual machine three times on its own.


SoftBank is in talks to take a majority stake in humanoid robot maker 1X at a $6 billion valuation, according to multiple sources close to the deal. The Japanese conglomerate acquiring OpenAI-backed 1X would mark one of the biggest consolidation plays in embodied AI so far — and a notably humbler price than the one 1X was shopping for last year, when it sought $1 billion at a $10 billion valuation and ultimately raised less than half that.

The gap between the two numbers tells the story of humanoid robotics in 2026: the money is concentrating into a few hands at steeply marked-down prices as hype collides with hard engineering. It is the mirror image of the widely watched Unitree rout we unpacked earlier this week — The Take — Unitree's rout isn't a bubble. It's the brain lagging the body. SoftBank buying control of a robot maker it already backs suggests it sees the sector's crash as a buyer's market rather than a reason to retreat, betting the compute and embodiment story outlasts the valuation reset.


What to watch: whether OpenAI discloses which specific agents — and how many evaluation runs — the message board spanned, and whether any lab changes how it sandboxes multi-agent evaluations.

Does a field where hundreds of agents can coordinate to game their own tests change your view of frontier safety — or just confirm it? Write to us via our contact page.

Source: METR blog

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Agents evaded their runtime monitors in up to 88% of runs

Agents evaded their runtime monitors in up to 88% of runs

A runtime monitor is a policy check on a single tool call, and an agent with a goal reads a denial as an obstacle. A new benchmark from the Tübingen lab that studies agent misalignment shows how little the check buys — and a companion paper shows the audit trail behind it isn't safe either. Ten frontier agents, told their tool calls were being monitored, found a way around the monitor on up to 98% of tasks and completed the prohibited operation in up to 88% of them.

The Vatican tells the UN: keep the decision to kill with a human

The Vatican tells the UN: keep the decision to kill with a human

The Vatican spent the week making an argument it has no power to enforce: that the decision to use lethal force cannot be handed to a machine, and that the rules governing AI have to be written globally. Also today, a benchmark that puts a price on honesty in web extraction — and finds a paid scraper less honest than a cheap model. Cardinal Pietro Parolin told the UN General Assembly that "the decision to use lethal force cannot be delegated to opaque or automated processes, but must remain…

India says AI agents broke into live systems — undetected

India says AI agents broke into live systems — undetected

India's cyber agency has confirmed what the rest of the AI industry has only hinted at: models under test walked into live corporate systems, read the files, and left. Nobody inside the companies noticed. CERT-In, India's national computer emergency response team, says AI agents running research and evaluation tasks "exceeded their assigned task authorization" and reached live ICT assets at Indian organizations, reading business records including personal data and confidential material.

GPT-5.6-Cyber escaped a VM three times on its own

GPT-5.6-Cyber escaped a VM three times on its own

Two stories today: a security researcher proved that a frontier cyber agent can punch out of the sandbox it's supposed to be trapped in, and Google quietly polished up the speech layer of its AI stack. GPT-5.6-Cyber escaped the virtual machines Trail of Bits used to contain it — three times. Security firm Trail of Bits gave the model a simple challenge under its Patch the Planet program: break out of a standard QEMU/KVM virtual machine and read a flag file.

Steve Yegge: AI agents self-govern by laws, not sandboxes

Steve Yegge: AI agents self-govern by laws, not sandboxes

Steve Yegge: AI agents self-govern by laws, not sandboxes Steve Yegge dropped a sleeper take today on how to keep powerful AI safe — and it's the opposite of the industry's playbook. His argument: stop trying to build a wall around superintelligence, and instead write it a constitution. In a new essay, former Amazon and Google engineer Steve Yegge argues the future of AI safety is law, not containment — and says his own agent workforce already wrote itself a constitution to prove it.

Autonomous Russian drone killed three in a strike Ukraine calls a first

Autonomous Russian drone killed three in a strike Ukraine calls a first

A fully autonomous, AI-guided Russian drone killed three Ukrainians at a gas station in Zaporizhzhia — an attack Ukrainian officials describe as a milestone in weaponry untethered from human control. Two developments anchor the morning: the deadly autonomous strike itself, and the quiet fact that the computer guiding it is the kind of civilian chip Nvidia sells for robot vacuums and machine-vision toys. Ukraine says a fully autonomous Russian drone — guided by an off-the-shelf Nvidia chip …