Security
An attack on world models can give robots dangerous policies
The authors put malicious prompts in data that looks safe, and the attack works only after a world model uses the data.
arxiv.orgClaimed, not confirmed
News category
Security
268 stories, newest first.
Security
A guardrails proxy increases injection recall from 25.0% to 94.5%
The authors found a false-positive rate of 0.1%, and the system limits the model to a closed set of commands.
arxiv.orgClaimed, not confirmed
Security
A new attack agent goes around the guardrails of 4 AI models
The researchers find a success rate of more than 95% for the attack.
arxiv.orgClaimed, not confirmed
Security
The WebMirage attack hijacks web agents in 91.9% of tests
The attack can also get around 3 agent-level defenses.
arxiv.orgClaimed, not confirmed