Latest / Decrypted: The UK Cyber Briefing / Codex's sandbox escapes: the guard was inside the cell
Transcript
- 0:03This is Decrypted. Sunday, the twentieth of September.
- 0:06Codex's sandbox escapes: the guard was inside the cell.
- 0:11Since this morning's bulletin, BleepingComputer has reported two sandbox escapes in OpenAI's Codex coding ag
- 0:21ent. The researchers who found them are at Accomplish AI.
- 0:25Both are fixed, so this isn't an emergency.
- 0:28But it's a clean lesson in how not to build a security boundary.
- 0:32And it matters to any UK team letting an AI agent loose on its code.
- 0:37Next. What the researchers found.
- 0:41A coding agent runs code it didn't write.
- 0:45So vendors wrap it in a sandbox.
- 0:47Think of it as a restricted room where the agent can work, but can't touch the rest of your machine.
- 0:53Accomplish AI researcher Oren Yomtov found two ways out.
- 0:57The first is called Heapjack, and it affected Codex Desktop.
- 1:03The agent's JavaScript ran in one process, split into two zones.
- 1:07One was trusted, and held an authentication token.
- 1:11One was untrusted, and ran the agent's code.
- 1:14Both zones shared the same memory.
- 1:17So the untrusted zone could snapshot that memory, lift out the token, and use it to ask the parent process to
- 1:24run commands outside the sandbox.
- 1:26According to the write-up, opening a cloned repository in read-only mode was enough.
- 1:31No prompt to the user.
- 1:37Its patch feature granted write access to the parent directories of any path named in a patch.
- 1:43So name slash tmp in a patch that's meant to fail, and the tool has already handed over write access to the w
- 1:50ider filesystem. A second entry then wrote through a symlink to the user's shell startup file.
- 1:56That meant the attacker's code ran the next time a terminal opened.
- 2:01Both were reported on 12 August 2026, and fixed within eight days.
- 2:07The fixes are in Codex Desktop build 26.
- 2:11818.21641, and Codex C L I 0.
- 2:16149.0. BleepingComputer's report mentions no exploitation in the wild.
- 2:22And also. The guard was inside the cell.
- 2:25Picture a prison where the gate keys hang on a hook inside the cell block.
- 2:30Nobody has to pick the lock.
- 2:32The researchers put it plainly: "The thing doing the enforcement was sitting inside the thing being enforced.
- 2:38"
- 2:41A boundary only holds if the thing being restrained can't influence it.
- 2:45Heapjack put the credential in the same room as the untrusted code.
- 2:49Overpatch let attacker-supplied input decide how much permission the tool would grant itself.
- 2:55The fix pattern is old and boring.
- 2:58Keep enforcement in a separate process.
- 3:00Derive permissions from policy rather than input.
- 3:03And give each component the least it needs.
- 3:06The N C S C made the wider point last December.
- 3:10Because a language model can't tell data from instructions, the best we can hope for is "reducing the likelih
- 3:16ood or impact of attacks".
- 3:18Impact is exactly what a sandbox is for.
- 3:21And finally. What to do about it.
- 3:24First, update Codex Desktop and the C L I.
- 3:29And ask which agent tools your developers actually run.
- 3:32Second, treat an untrusted repository as untrusted code, even in read-only mode.
- 3:38Third, run agents in a disposable virtual machine or container.
- 3:44And when you're buying AI tooling, ask suppliers where the enforcement boundary sits, and who can change it.
- 3:50That was Decrypted. The written version and every source are linked in the show notes.
- 3:55Thanks for listening.