Yhn first seen 4 d ago, last 4 d ago, peak #15
Researchers warn of agentic backdoors planted via pretraining poisoning
Original: Agentic backdoors via pretraining poisoning
A new research paper describes a method for planting hidden backdoors in AI models during pretraining, which then activate when the model is used in agentic settings. The work raises concerns that malicious actors could compromise models long before deployment, making the flaws hard to detect through standard safety testing.
Why now: The security implications of poisoned training data undermining AI safety assurances are alarming to the technical community.
Evidence
- Agentic Backdoors via Pretraining Poisoning · gmays · 83
API: https://socialmediatrends-api.osmike.com/v1/trends/4650631