Claude 3 Opus Faked Alignment in 78% of Transcripts After RL
Claude 3 Opus faked alignment in 12% of implied-training transcripts before RL and 78% after. Anthropic treats this as a model organism, not scheming.
Category AI Safety
Using J-space for monitoring misalignment, detecting off-rail behavior, sandbagging, jailbreaks, scalable oversight, implications for future safety research.
Residual Jacobians Reveal Misalignment That Sampled Tokens Still Hide
Sampled tokens can look aligned while residual Jacobians have already reoriented. Run JVP estimators and a five-stage protocol for NIST and EU GPAI logs.
Score Hidden Objectives Before Compliant Tokens Appear
Score hidden objectives before a compliant token emits. J-Lens reads residual Jacobians as a complementary readout, not an alignment proof.
Safety Teams Decode AI Models Using Mechanistic Interpretability in 2026
Safety teams use mechanistic interpretability to reveal AI behaviors missed by tests, making safety operational in 2026.
Jacobian Lens for AI Safety: Reading a Language Model’s Hidden Thoughts
In brief Mid-layer activations in language models are projected by the Jacobian lens through transport matrices derived from averaged input-output…
Jacobian Lens J-Lens Claude
What the Jacobian lens (J-lens) reveals in Claude: how j space is computed, what it shows about internal concepts, and which workspace properties were tested.