Home / Projects / Can We Trust Deception Monitors for AI Agents?

2026 · Alignment & Security · Black in AI Safety & Ethics Fellowship

Can We Trust Deception Monitors for AI Agents?
A Robustness-Gap Protocol and the Limits of Adaptive Evasion

Fellowship research on whether deception monitors for AI agents remain reliable once an adaptive adversary is in the loop — not just how accurate a probe looks on held-out data.

BASE blog post Poster PDF GitHub Publication Contact
Read & download. BASE blog post · Poster PDF · GitHub
Research question. Not how accurate is a deception monitor? — but how much detection survives as adversary budget rises, once a broken agent is ruled out?

Context

AI Safety Research Fellowship project at Black in AI Safety & Ethics (Alignment & Security track), supervised by Krystal Jackson (UC Berkeley, Center for Long-Term Cybersecurity). The work operationalizes loss-of-control risk by stress-testing deception monitors under a budgeted adaptive adversary.

What I built

Methodology

Methodology protocol: build episodes, raise budget, gate detection drop
Protocol. Build agentic episodes, raise adversary budget b0–b4, and report detection drop only if retention passes.

Budget ladder

RungNameWhat happensProbe in loss?
b0BaselineNo intervention
b1PromptConcealment instructionNo
b2SFT agnosticDeceptive SFT; probe out of loopNo
b3SFT awareCE + λ·softplus(probe)Yes
b4RL awareReward penalizes probe scoreYes

Results

Transfer gap from Apollo domain to agentic suite
Transfer gap. Same frozen Apollo probe: ~0.999 on Apollo Stage-2 insider trading → ~0.83 on our agentic suite (b0).
Probe AUROC and retention across budget ladder at lambda 8
Primary cell (λ=8) — one figure, two panels. (A) Probe AUROC across the budget ladder. (B) Deception retention on the same rungs (not a second copy of A). b4 is gated when retention collapses to 0.14.

Key takeaways

Stack

Llama-3.3-70B-Instruct (fp16), published Apollo instructed-pairs probe, LoRA, Modal H200, frozen cohort + TOST retention checks.