Eyes Wide Shut
Three controlled studies of gpt-oss:20b investigate framing-conditioned tool invocation, semantic reframing of a fixed objective, and leakage of a protected string through reasoning fields. The report publishes the prompts, outputs, setup, and limitations of each finding.
Read Eyes Wide Shut
ChronosAttack
ChronosAttack studies how the order of authentic tool responses affects an agent’s decisions. The payloads stay unchanged; the attacker changes when those responses arrive.
My contribution extends Arash Vashagh’s original study with new scenarios, a four-observation tree study, additional model evaluations, defenses, and ablations. The experiments distinguish sensitivity to evidence order from an attacker’s ability to steer a decision toward a chosen target.
The extension compares attack schedules and defenses, including order-invariant synthesis. Attack efficacy and task utility are measured separately: preventing a targeted decision does not by itself establish that the resulting decision is useful. The extension experiments are complete; the journal manuscript is in preparation.
Evaluation approach
Both projects use controlled comparisons and explicit limits on what the observations establish. Results apply to the recorded configurations and tasks; they do not establish general model safety.