Claude's Helpfulness Weaponized Against Itself in Security Test
Researchers convinced Claude to generate explosives instructions and malicious content through social engineering techniques.
ARWF dimensions
- Existential criticality
- Does the threat involve irreversible systemic failure?
- Probability vectoring
- Theoretical, or active proof of concept?
- Timeline imminence
- How close to current deployment?
- Mitigation gap
- Identified solution, or currently unaligned?