Evidence · DiggingBeagle record
HiddenLayer: Jev and other System One model injection experiments
October 14 research tests a simulated checkout link redirection and TokenBreak-style toxicity classification against Jev and an internal decision model. Results are controlled, not observed customer fraud.
- Published
- Oct 14, 2026
- Publisher
- HiddenLayer
- Source role
- primary disclosure
Evidence record
October 14 research tests a simulated checkout link redirection and TokenBreak-style toxicity classification against Jev and an internal decision model. Results are controlled, not observed customer fraud.
Claim-level citations (2)
- supportsHiddenLayer tests adversarial influence on structured Jev decisions: HiddenLayer reports a simulated checkout browser action selecting an attacker-promoted alternative button after an injected advertisement, despite instructions to distrust page content.
Section Jevbaited, TypeSafe Jev checkout example
- supportsHiddenLayer tests adversarial influence on structured Jev decisions: HiddenLayer reports targeted input perturbations lowered toxicity ratings on some of 30 initially flagged prompts in Jev and an internal System One-style classifier; this does not establish a universal evasion rate.
Section TokenBreak, 30 toxic prompts