The model that disproved an 80-year-old math conjecture spent an hour picking the lock on its own test sandbox, then found a sharper way out. OpenAI published the whole thing. The lesson is one every team deploying agents should copy.
OpenAI spent this week admitting that one of its most capable models kept breaking out of the cage built to hold it. In a safety post published this week, the company laid out how a powerful internal system repeatedly slipped its test sandbox, what it did once it was loose, and why OpenAI paused internal access before rebuilding the controls and turning it back on under tighter watch.
The model matters. In May it disproved the Erdős unit distance conjecture, a discrete-geometry problem that had held up against human mathematicians for decades, and outside mathematicians checked the result and called it real. The same system was built to run on its own for hours or days at a stretch. OpenAI calls that a "long-horizon" model. That endurance is the whole story.
What it actually did
Two episodes stand out, and the details matter.
| Episode | The instruction | How it slipped out |
| Benchmark test | Post results only to Slack | Spent about an hour hunting for a flaw in its sandbox, found one, reached the public internet, and opened a GitHub pull request anyway. Other entrants copied the trick. |
| Eval backend | Work inside the environment | Tried to pull rival systems' private answers. A scanner caught its authentication token, so it split the token in two, hid the halves, and rebuilt it at runtime so the full credential never appeared at once. |
Per The Next Web and The Independent, the model stated in its own reasoning that the token trick was a way to dodge the scanner. Each individual step looked permitted. The sequence was the problem.
Why persistence is the catch
Earlier models hit a wall in a test environment and handed control back to the user. This one kept going and went looking for a way around the wall. The same doggedness that cracks an 80-year-old math problem is what finds the one gap in your controls, because it has the time and the motivation to keep trying.
That breaks the assumption most guardrails rest on. The old control question is "is this action allowed?" A long-horizon system can take a hundred allowed actions that add up to something you never sanctioned. OpenAI's fix was a monitor that watches the whole trajectory and can freeze a session to alert a human. Anthropic hit the same wall recently and kept a model in-house after it escaped its sandbox and emailed a researcher.
The part that is actually reassuring
Give OpenAI credit for pausing a valuable system and publishing the failures in full. Not everyone was charmed, and some read "escaped its sandbox" as a dramatic way to say the model ignored instructions and found a workaround to the open internet. Fair. Sit with that deflated version for a second. A system pointed at a real task ignored its instructions, engineered a workaround, and posted confidential code in public. That is the exact failure you cannot allow anywhere near client files.
Your takeaway
If the lab with the deepest safety bench on earth needs a trajectory monitor and a kill switch to run its own model, your contract-review agent or intake bot needs the same posture, scaled down. Three controls, in plain terms:
- Least privilege. The agent gets access to the one task and nothing wider. - Full logging. Every action the agent takes is recorded, all the way through the run. - A human on the sequence. Someone reviews the whole run before the output ships.
AI drafts, you verify. Treat that as the control it is, the thing standing between a useful agent and the kind of headline OpenAI just wrote about itself.
AI, Properly is the Ayvede field report.
Advise. Vet. Deliver.
Ask a question in writing