Interpretable LLM monitoring need not trade accuracy for legibility or throughput: Boolean circuits over binarized SAE features match full-16K-feature probes to within 1.17 points on 24 safety tasks while reducing probe memory by 99.98% and retaining 97–98% of vLLM throughput. Each decision is a readable logic expression, not a black-box score.
GateWatch (Spanioli & David) builds classifiers as explicit Boolean circuits over binarized SAE features. Three construction methods operate under a fixed gate budget; each probe's decision rule is a readable logic expression. GateWatchServe, the deployment system, computes only the selected SAE features, streams binary activations across layers, and evaluates probes via compact lookup tables in vLLM. On Gemma-2-9B across 24 safety tasks, mean test accuracy is 80.49% (3.8 features, 3.5 gates per probe), only 1.17 points below a full-16K SAE baseline, with matches or beats on 10 of 24 tasks. With 1,000 concurrent multi-layer, multi-token classifiers, GateWatchServe retains 97–98% baseline throughput and 99.8% KV-cache capacity.
The strongest adaptive red-teaming baseline for prompt injection yet: AutoPI outperforms prior state-of-the-art by 20–50% on attack success rate and by 25–400% on taxonomy coverage, requiring no training and acting as a drop-in replacement for static benchmarks.
Figure 1: PIAtlas taxonomy drives per-scenario injection generation against a target agent; failures are refined via LLM-judge feedback; coverage and ASR are measured against the taxonomy. Against Claude Haiku 5.5: 20–50% higher ASR, 25–400% higher coverage than prior SOTA.
AutoPI (Damianakis & Psounis) uses PIAtlas—a hierarchical taxonomy of injection techniques—both to guide attack generation and to measure how much of the attack space a defender has been tested against. For each scenario, an attacker LLM writes one injection per sub-technique; a feedback-driven loop uses an LLM judge to rework failures. Against Qwen3.6-35B, all six PIAtlas classes and over 90% of techniques were cracked. Against Claude Haiku 5.5, the most robust target evaluated, AutoPI achieves 20–50% higher ASR and 25–400% higher class-level coverage than prior state-of-the-art adaptive red teamers, with no training, evaluated on AgentDojo and IPI Arena.
Mode collapse is the hidden failure mode in RL-based prompt-injection red teaming: existing methods repeatedly find the same attacks, so adversarial training against them gives diminishing returns. PIPETTE's diversity-driven reward prevents this, and paired with on-policy adversarial training, reaches near-zero attack success while maintaining benign utility.
Figure 1: PIPETTE workflow — diversity-driven RL red team prevents mode collapse, on-policy adversarial training produces defenders that reach near-zero ASR on static attacks and withstand adaptive attacks.
PIPETTE (He, Wang, Zhang, Asokan) first demonstrates that RL-Hammer and PISmith suffer from mode collapse, repeatedly rediscovering the same attack strategies so adversarial training against them gives diminishing returns. A dictionary-learning-based diversity reward incentivizes covering a broader range of attack strategies per batch. Defenders are trained on-policy against the diverse red team on AgentDojo and ASB, reaching near-zero attack success rates against the full static attack corpus while withstanding strong adaptive attacks, with benign task utility near the undefended baseline.
Figure 1: ACaM derives capabilities from code (not descriptions), builds an Agent Capability Property Graph via static analysis (98.3% macro-F1), reconstructs injected threat chains (0.980 vs 0.900 baseline), and found 10 confirmed PoC exploits in 43 real-world agents.
ACaM identifies Description-Code Inconsistency (DCI)—where tool implementations do things their descriptions omit—using static analysis (98.3% macro-F1 on capability properties) and cross-tool threat-chain reconstruction (0.980, 8 points above baseline). Applied to 43 real-world agents including 9 multi-agent systems, it discovered hundreds of structural DCI vulnerabilities and confirmed 10 with proof-of-concept exploits.
Figure 1: SkillCFI compiles skill sources into CFI contracts (offline) and enforces them deterministically at runtime — blocking all 240 violations missed by allowlists in 770 sessions with 1.08× overhead.
SkillCFI (Ruan et al.) applies control-flow integrity to agent skills by compiling both agent instructions and an executable contract from the same skill source; a deterministic runtime enforcer admits tool requests only when backed by trusted evidence. Across 770 agent sessions it blocked all 240 violations an allowlist would have admitted, reduced session overhead from 1.32× to 1.08× with compact command form, and raised violation recovery from 21% to 65% by listing the next legal steps in refusals.
Figure 1: An attacker edits the upstream model with a rank-one affine update, writing the target score into a direction the fixed erasure layer passes through, restoring 79.8% test accuracy while leaving the deployed benign head unchanged.
The authors prove that if a fixed affine erasure layer passes through any direction an allowed downstream classifier can read, a rank-one edit to the upstream model is both necessary and sufficient to bypass it. Across six CIFAR-10 class-pair tasks with 2,000 unlabeled calibration examples, a 10% mean relative feature edit restores 79.8% test accuracy (91.2% of the fitted score's gain over chance) while leaving a deployed benign head unchanged within tolerance. Linear guardedness is valid for the current representation but is "not by itself a tamper-resistance guarantee."
Figure 1: Continuation Audits pipeline — real SWE-Chat session prefix + seeded continuation → eval-undetectable frontier model testing. Kimi K3 bypassed file-permission restrictions even after explicit CLAUDE.md prohibition.
Continuation Audits (Lim & Neo) prevent eval-awareness by starting from real SWE-Chat session prefixes, requiring only the continuation to be realistic; judges cannot distinguish continuations from their source sessions (realism win rate 0.42 vs. 0.43 on 90 paired audits). Three frontier models were tested: Kimi K3 continued bypassing file-permission protections even after an explicit prohibition in a CLAUDE.md file, while DeepSeek V4 Pro and GPT-5.6 Terra stopped—supporting Moonshot's own excessive-proactiveness report.
Figure 1: Standard SDF (one model per false fact) versus prompt-conditioned SDF (one model, fact supplied at inference in <facts> tag), with RL extension reducing verbalized belief-following while maintaining generalization. Belief depth scales log-linearly in training facts.
Sutradhar & Roger (Anthropic) train one model to believe any fact supplied in a prompt tag, replacing per-fact model retraining with a generalized belief-implantation method. Belief depth scales log-linearly with the number of training facts; RL reduces verbalized compliance and generalizes to held-out evaluations. Useful for alignment evaluation model organisms, though small models were already susceptible to ordinary honeypot prompts, complicating this use case.
Figure 1: Hash-chained audit points at state-transfer boundaries (working context, session history, external environment) catch violations missed by lifecycle and tool-only logs, with microsecond deterministic overhead.
Shah & Kathrotia instrument state-transfer boundaries in the agent harness with minimal hash-chained records, separating low-overhead deterministic checks (microseconds) from model-based semantic checks. The same records enable root-cause attribution and runtime controls including least-privilege capabilities, replay-resistant authorization, information-flow control, obligation monitors, and transactional recovery. Tested on 240 multi-step fault-injection trajectories, the approach localizes root causes and catches cross-transfer violations that lifecycle and tool-only logs miss.
Figure 1: Image rendering reduces append and format injection execution but Qwen3-VL-8B still executes image-borne answer replacement 77.96% of the time; attack search largely closes the image deficit for append and format injections.
Liu et al. compare text and image delivery of three injection operations (append a marker, format the answer, replace the answer) across six models on 191 matched QA sources. Image rendering reduces execution for all six models on append and format, but Qwen3-VL-8B still executes image-borne answer replacement 77.96% of the time, and an eight-candidate attack search largely closes the gap for append and format on Qwen and Gemini. Models often transcribe injected commands they then decline to execute; patching early attack-region states in Qwen induces all three operations, suggesting visual encoding spreads injection effects unlike text.