WEBVTT

1
00:00:03.200 --> 00:00:06.880
This is Decrypted. Sunday, the twentieth of September.

2
00:00:06.880 --> 00:00:11.920
Codex's sandbox escapes: the guard was inside the cell.

3
00:00:11.920 --> 00:00:21.040
Since this morning's bulletin, BleepingComputer has reported two sandbox escapes in OpenAI's Codex coding ag

4
00:00:21.040 --> 00:00:25.280
ent. The researchers who found them are at Accomplish AI.

5
00:00:25.280 --> 00:00:28.880
Both are fixed, so this isn't an emergency.

6
00:00:28.880 --> 00:00:32.720
But it's a clean lesson in how not to build a security boundary.

7
00:00:32.720 --> 00:00:37.600
And it matters to any UK team letting an AI agent loose on its code.

8
00:00:37.600 --> 00:00:41.520
Next. What the researchers found.

9
00:00:41.520 --> 00:00:45.040
A coding agent runs code it didn't write.

10
00:00:45.040 --> 00:00:47.520
So vendors wrap it in a sandbox.

11
00:00:47.520 --> 00:00:53.200
Think of it as a restricted room where the agent can work, but can't touch the rest of your machine.

12
00:00:53.200 --> 00:00:57.920
Accomplish AI researcher Oren Yomtov found two ways out.

13
00:00:57.920 --> 00:01:03.360
The first is called Heapjack, and it affected Codex Desktop.

14
00:01:03.360 --> 00:01:07.840
The agent's JavaScript ran in one process, split into two zones.

15
00:01:07.840 --> 00:01:11.440
One was trusted, and held an authentication token.

16
00:01:11.440 --> 00:01:14.960
One was untrusted, and ran the agent's code.

17
00:01:14.960 --> 00:01:17.600
Both zones shared the same memory.

18
00:01:17.600 --> 00:01:24.320
So the untrusted zone could snapshot that memory, lift out the token, and use it to ask the parent process to

19
00:01:24.320 --> 00:01:26.400
run commands outside the sandbox.

20
00:01:26.400 --> 00:01:31.600
According to the write-up, opening a cloned repository in read-only mode was enough.

21
00:01:31.600 --> 00:01:37.600
No prompt to the user.

The second is Overpatch, and it affected the command-line tool.

22
00:01:37.600 --> 00:01:43.600
Its patch feature granted write access to the parent directories of any path named in a patch.

23
00:01:43.600 --> 00:01:50.747
So name slash tmp in a patch that's meant to fail, and the tool has already handed over write access to the w

24
00:01:50.747 --> 00:01:56.960
ider filesystem. A second entry then wrote through a symlink to the user's shell startup file.

25
00:01:56.960 --> 00:02:01.160
That meant the attacker's code ran the next time a terminal opened.

26
00:02:01.160 --> 00:02:07.120
Both were reported on 12 August 2026, and fixed within eight days.

27
00:02:07.120 --> 00:02:11.440
The fixes are in Codex Desktop build 26.

28
00:02:11.440 --> 00:02:16.800
818.21641, and Codex C L I 0.

29
00:02:16.800 --> 00:02:22.720
149.0. BleepingComputer's report mentions no exploitation in the wild.

30
00:02:22.720 --> 00:02:25.920
And also. The guard was inside the cell.

31
00:02:25.920 --> 00:02:30.400
Picture a prison where the gate keys hang on a hook inside the cell block.

32
00:02:30.400 --> 00:02:32.480
Nobody has to pick the lock.

33
00:02:32.480 --> 00:02:38.520
The researchers put it plainly: "The thing doing the enforcement was sitting inside the thing being enforced.

34
00:02:38.520 --> 00:02:41.120
"

That's the Secure by Design point.

35
00:02:41.120 --> 00:02:45.520
A boundary only holds if the thing being restrained can't influence it.

36
00:02:45.520 --> 00:02:49.840
Heapjack put the credential in the same room as the untrusted code.

37
00:02:49.840 --> 00:02:55.680
Overpatch let attacker-supplied input decide how much permission the tool would grant itself.

38
00:02:55.680 --> 00:02:58.320
The fix pattern is old and boring.

39
00:02:58.320 --> 00:03:00.800
Keep enforcement in a separate process.

40
00:03:00.800 --> 00:03:03.760
Derive permissions from policy rather than input.

41
00:03:03.760 --> 00:03:06.400
And give each component the least it needs.

42
00:03:06.400 --> 00:03:10.000
The N C S C made the wider point last December.

43
00:03:10.000 --> 00:03:16.865
Because a language model can't tell data from instructions, the best we can hope for is "reducing the likelih

44
00:03:16.865 --> 00:03:18.640
ood or impact of attacks".

45
00:03:18.640 --> 00:03:21.760
Impact is exactly what a sandbox is for.

46
00:03:21.760 --> 00:03:24.800
And finally. What to do about it.

47
00:03:24.800 --> 00:03:29.120
First, update Codex Desktop and the C L I.

48
00:03:29.120 --> 00:03:32.160
And ask which agent tools your developers actually run.

49
00:03:32.160 --> 00:03:38.800
Second, treat an untrusted repository as untrusted code, even in read-only mode.

50
00:03:38.800 --> 00:03:44.040
Third, run agents in a disposable virtual machine or container.

51
00:03:44.040 --> 00:03:50.240
And when you're buying AI tooling, ask suppliers where the enforcement boundary sits, and who can change it.

52
00:03:50.240 --> 00:03:55.040
That was Decrypted. The written version and every source are linked in the show notes.

53
00:03:55.040 --> 00:03:56.400
Thanks for listening.

