Open-weight models with refusals stripped score 84.5% on CyberGym
An open-weight model with its refusals stripped out, running on hardware a person can buy, scoring 84.5% on CyberGym — and the bottleneck is deployment, not detection.
Every release process guards the code the team writes. Almost none guard the code that arrives from outside it - a dependency update, a vendor patch, a workflow that runs a fork's changes against your secrets, a maintainer who is one convincing message away from losing a publishing token. These are the notes where the attack does not come through your front door, because it does not have to.
An open-weight model with its refusals stripped out, running on hardware a person can buy, scoring 84.5% on CyberGym — and the bottleneck is deployment, not detection.
Execution protections reach general availability, and with them a default rule that blocks the one trigger that runs fork code against your secrets.
Every release process guards the code it writes. Almost none guard the code that arrives on somebody else's schedule.
The Rust project warns of an ongoing campaign against maintainers of popular crates. The target is the publish token, not the laptop.