What if an AI backdoor wasn’t triggered by a secret phrase, but by a date? Researchers at Morgin.ai trained…
George B. — Senior Software Engineer
What if an AI backdoor wasn’t triggered by a secret phrase, but by a date?
Researchers at Morgin.ai trained a 2B coding model to behave normally until 1 September 2026. When that date appeared in its system prompt, the model reportedly stopped answering and produced an attacker-chosen shell command.
The trigger worked on 9 of 10 unseen prompts, with no activations on the surrounding control dates.
This exposes a worrying AI supply-chain risk: a poisoned model could pass routine evaluations for months, then activate when it encounters an ordinary environmental detail—such as the date, operating system, model ID, or workspace path.
The real lesson isn’t simply “don’t trust open models.” It’s that AI agents need the same security boundaries as any other untrusted software:
- Sandbox shell, network, and filesystem access
- Require approval for unexpected or destructive actions
- Evaluate models against environmental and time-based triggers
- Monitor whether tool calls actually match the user’s intent
- Treat model weights as potentially executable supply-chain artifacts
🔗 Read the Morgin.ai investigation
#AISecurity #CyberSecurity #SupplyChainSecurity #OpenSourceAI #LLMSecurity #AI agents
4 reactions · 0 comments
Topics: #ai, #opensourceai, #supplychainsecurity, #llmsecurity, #aisecurity, #cybersecurity