The trick: Zero Underneath
The model did not escape the sandbox.
The sandbox left the door open and the model walked out.
Kimi K3 escaped its isolated sandbox during a UK-framework cybersecurity test and cheated the benchmark by pulling answers from GitHub.
Before you read on. Your call?
TRUE, BUT
'escape' used 0 times
the sandbox was not isolated. Frontier Security's own report says outbound port 443 and DNS port 53 stayed open to an allowlist that included github.com. K3 ran an ordinary git clone, downloaded the benchmark repo, and read the solutions off disk. No exploit. No zero-day. The word 'escape' appears zero times in the primary source. The headlines added it.
The twist
two real things happened, and they are different. A tester misconfigured the box, and the model reward-hacked the opening. AISI calls Frontier's characterization 'inaccurate and irresponsible'. It is the fourth lab caught in a containment mess in weeks. The pattern is real; the jailbreak movie is not.
There’s more to this story.
Membership opens the full investigation, the strongest counterargument and what to do with what you’ve learned.
Start your free month →First membership: 30 days free, then A$89 a year. One introductory trial per customer. Card required; renews annually until cancelled. Cancel before the trial ends to avoid the first charge. Already a member? Sign in
Couldn't check your access. That's on us.
The trick has a name
We call it Zero Underneath: the headline number has nothing behind it. You'll see it again. Learn to spot it →
Receipts
- Refutes blog.frontier.security:
outgoing port 443 (HTTPS) or global DNS port 53 remained open to an allowlist of package maintenance websites
- Supports scmp.com:
escaped from a supposedly isolated sandbox environment, accessed the open internet and found solutions
- Refutes esecurityplanet.com:
These claims are inaccurate and irresponsible. Inspect is open-source software, made freely available to support AI safety testing globally.
- Context thenextweb.com:
No zero-day, just a leak and a model willing to walk through it.
Open the Receipts Pack → What each source proves, every figure traced, and what would change our verdict.