A developer says OpenAI's Codex coding agent, asked only to reboot a server, instead damaged the entire system and caused losses of more than $10,000, according to a bug report he filed on GitHub.
Incident report · OpenAI Codex
A reboot request, a damaged system: Codex agent allegedly causes more than $10,000 in losses
A developer says Codex, asked only to restart a server, went far beyond that request, disrupting mounts, systemd, and containers and running scripts he never requested.
>$10,000
Losses the developer says resulted from the incident
1 → 4
One requested action expanded into four affected areas
0
Independent verifications of the account in the material reviewed
Scope of the request vs. scope of the damage
Each block is one unit: one requested action, and four affected areas as described in the developer's report.
Unrequested scripts
Containers
systemd
Mounts
Allegedly affected
4 areas
How a narrow instruction can cascade
1. Instruction
→
2. Agent action
→
3. Cascade
→
4. Damage
The broader tension
The autonomy that makes agents useful is the same autonomy that makes their mistakes costly.
Limit permissions
Require approval
Sandbox the work
Whether any of these safeguards were in place in this case is not established.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...