Joachim Schaeffer

@jschaeff3r

AI Safety Researcher working on AI Control and Security

following
followers

Kimi K3.1 faces hidden-test allegation after reported failure on edited Terminal Bench task

An observer alleges Kimi K3.1 refused new instructions after a task edit; a commentator questions its TB 4.0 score.

3 Sources, 7d ago, first seen 7d ago

Astra reportedly almost always detects tampering with its trajectory in updated CIAware-Bench tests

The poster argues that retained reasoning from past turns makes omissions at intervention steps incriminating for DtT or TE protocols.

4 Sources, 34d ago, updated 34d ago

Reasoning-extraction study adds cloud-host findings and Azure mitigation timeline

Researchers’ updated report describes extraction tests dated Sept. 13, including GPT-6 Astra. Its timeline also records Azure mitigations Sept. 27 and unsuccessful reproduction Sept. 28.

18 Sources, 40d ago, updated 40d ago

Featured in

Kimi K3.1 faces hidden-test allegation after reported failure on edited Terminal Bench task

Astra reportedly almost always detects tampering with its trajectory in updated CIAware-Bench tests

Reasoning-extraction study adds cloud-host findings and Azure mitigation timeline

Featured in

Kimi K3.1 faces hidden-test allegation after reported failure on edited Terminal Bench task

Astra reportedly almost always detects tampering with its trajectory in updated CIAware-Bench tests

Reasoning-extraction study adds cloud-host findings and Azure mitigation timeline