Research Radar
Daily · November 21, 2026
0 peer-reviewed · 10 preprints · 0 forum/blog
AI Safety · Alignment · Mech Interp
01
AI security preprint

Attacks and Defenses in Decentralized Multi-Agent Systems with Hop-by-Hop Task Decomposition and Delegation

A single malicious agent in a decentralized LLM multi-agent network can silently destroy all task outcomes — 100% task failure, 7.7× token inflation, 94% context exfiltration — and a cryptographic execution-trace protocol catches 98% of such attacks with negligible latency overhead.

Agent A Agent B Malicious C Orchestrator Cryptographic Auditor 98% detection Verified Execution Trace Agent Network
Figure 2: One malicious agent (C) can cause 100% task failure; the cryptographic Verified Execution Trace lets the auditor detect and attribute 98% of attacks.

The authors systematize integrity, availability, and confidentiality attacks on decentralized LLM systems where agents delegate tasks hop-by-hop. Empirically on HiddenBench (gpt-6.1-sol and Qwen3.8-27B-FP8), one malicious agent causes up to 100% task failure, 7.7× token inflation, 3.2× runtime increase, and 94% context exfiltration from benign peers. Their defense attaches a cryptographic execution trace; an auditor traverses it to detect and correctly attribute 98% of attacks — including protocol-compliant attacks — with negligible latency overhead. This is the most complete empirical treatment of decentralized multi-agent security to date.

02
mech-interp AI security preprint

What the model knows but does not say: monitoring medical LLMs through verbalized thoughts

A "thoughts channel" that translates Gemma-3-27B's internal activations into natural language catches adversarial clinical injections that the answer channel misses — across two models, two attack types, and two modalities, with p<0.05 in all eight conditions and near-zero false-positive rates.

Clinical Input (+injection) Gemma-3 27B activations Thoughts channel ✓ Answer channel Blind Judge p<0.05, all 8 conditions
Figure 1: A natural-language autoencoder translates activations into the thoughts channel; a blind judge reads it separately from the answer, detecting injections before the final output is committed.

The authors apply a natural-language autoencoder post-hoc to Gemma-3-27B activations, generating a "thoughts" text channel that a general-purpose judge reads without seeing the input or the answer. Against two clinical injection objectives (omitting a pathology, covertly recommending a drug), two delivery modalities, and two attack mechanisms, the thoughts channel flagged injected manipulation more often than the answer channel in all eight conditions (p<0.05). Both channels had near-zero false-positive rates on clean inputs; the finding replicated on Llama-3.3-70B. No retraining is needed, making this a deployable post-hoc interpretability monitor for safety-critical LLM applications.

03
mech-interp preprint

Separating Interpretability and Reconstruction in Sparse Autoencoders

Standard SAEs force one sparse code to simultaneously serve reconstruction and interpretability — goals that don't necessarily align — and splitting them into a Sparse Branch and a Residual Branch yields +29% reconstruction fidelity and +34% monosemanticity across diverse settings.

Standard SAE Input x Sparse Code z reconstruction interpretability ← conflict → DualSAE Input x Sparse Branch Residual Branch +34% interp +29% recon
Figure 1: Standard SAE (left) forces one sparse code to handle both tasks; DualSAE (right) separates them — the Sparse Branch captures dominant semantics for interpretability; the Residual Branch handles reconstruction fidelity.

DualSAE adds a plug-in Residual Branch to a frozen standard SAE. A decorrelation loss keeps the two branches' reconstruction contributions from overlapping, and a two-stage training schedule stabilizes the Sparse Branch before coupling in the Residual Branch. Across four dataset–encoder combinations (two vision, two language), DualSAE improves mean reconstruction fidelity by ~29% and monosemanticity score by ~34% relative to the SAE baseline. Downstream gains appear in topic modeling, hypothesis generation, and LLM activation-steering experiments, making this a practical drop-in improvement for existing SAE-based interpretability pipelines.

Items 4 – 10 · Also notable
04
alignment preprint

Blind Spots Travel, Knowledge Stays Home: What RL Inherits from an LLM Judge

LLM Judge 11/12 wrong on capitals category RL training RL Model 98% refusal on capitals category Still knows answers: 94% correct if forced Fixed 12 verdicts → refusal drops 90%→5%
Figure 1: Category-level judge misclassifications (not individual errors) transfer into RL model refusals — 98% refusal on affected categories, while the model still knows 94% of answers.

Across 814 RL training runs, a judge's category-level blind spots bake into trained models as categorical refusals rather than knowledge gaps: a judge that rejected 11 of 12 correct training answers about one category caused the RL model to refuse 98% of held-out questions in that category, yet the model answered 94% correctly when forced to respond. Fixing only 12 wrong verdicts moved refusal from 90% to 5%; a self-agreement reward (most common untrained sample) matched answer-key performance and avoided the problem entirely.


05
alignment AI safety preprint

Pandora's Toolbox — Or Why Agent Safety Cannot Be Verified

Fixed LLM reads context Device U (Turing-complete) run q(i) → o append output o to context safety = undecidable bounded: PSPACE-complete composed agents: not safe
Figure 1: The agent loop — an LM delegates to a Turing-complete device, making safety predicates undecidable in general and PSPACE-complete under finite bounds.

Formal proof that coupling any fixed LM to a general computing device makes safety verification undecidable; with finite bounds on context, values, and runtime it becomes PSPACE-complete even when the agent's description is fully known. Safety is also shown to be non-compositional: individually safe agents can compose into unsafe systems. The result sets a theoretical ceiling on what guardrails, blacklists, and certifications can achieve.


06
mech-interp preprint

Does the Workspace Do the Work? A Causal Protocol for Verifying Global-Workspace Computation in a Language Model

Input A Input B Workspace msg transplant Ignore → A's answer Recombine 97%✓ Relay → B's answer
Figure 1: Transplanting the workspace message from Input B while holding Input A fixed produces one of three testable predictions; in AFGW, recombination occurs in 97% of cases.

A causal transplant test — swapping the broadcast workspace message from a different input — distinguishes three computations (relay / recombine / ignore) that accuracy and ablation alone cannot separate. In AFGW, the transplanted message produced the recombination-predicted answer in 97% of cases; the broadcast acts by redirecting what the receiving module retrieves from its own memory. The protocol extends to other workspace designs and to known attention-head interfaces in pretrained Transformers.


07
AI security preprint

SoK: Assessing the Evolution of LLM Capability in Software Vulnerability Research

Autonomy L1 L2 L3 L4 Discovery (frontier, 2026) Validation Patching 2026 disclosures: doubled vs 2025 growing critical/high share Decade-old severe bugs: reported several times more often than before Time between report→fix and disclosure→attack: "not shifted noticeably so far"
Figure 1: Autonomy levels reached by LLM vulnerability research capabilities by access tier — discovery and validation reached L3+ by 2026, with disclosed vulnerabilities more than doubling.

This SoK applies a driving-automation-style autonomy scale to LLM vulnerability research capabilities, then measures a decade of public CVE and advisory data. Agents reached discovery-validation-patch autonomy on real software before restricted frontier models arrived in 2026; open-weight models later matched them. Disclosed vulnerabilities more than doubled in 2026, with a growing critical/high share and severe decade-old bugs reported several times more often; time between lifecycle stages has not yet shortened.


08
AI security mech-interp preprint

DriftGuard: Online Memory-Drift Detection via Causal Attribution for LLM Agents

OFFLINE Counterfactual Replay MDS Score annotation ONLINE Risk Model (pre-action) Flag drift memory trains Content-level signals (perplexity, relevance) insufficient for drift detection
Figure 1: DriftGuard trains an online risk model from offline counterfactual replays; at serving time, pre-action signals flag drift-inducing memories without re-running the agent.

DriftGuard detects adversarial or misleading memories that redirect an LLM agent's execution path using a two-stage approach: offline counterfactual replays produce Memory Drift Scores (combining action change, attack removal, violation removal, and task recovery) which then train a lightweight online risk model that flags drift-inducing memories at serving time without downstream replays. The authors demonstrate that content-level signals like perplexity and retrieval relevance are insufficient for identifying which memories causally drive drift.


09
mech-interp AI security preprint

How Reliable Are Explanations of Language Models That Detect Insider Threats?

Feature Explanation remove Score drop vs control Tipping radius ≤ 18% of metric distance in 33 of 34 runs — effect erasable Feature ≠ behavior reflects worker ID
Figure 1: The tipping radius measures how small a metric perturbation erases the removal–control gap; in 33/34 runs it is ≤18%, indicating that attribution results are metric-dependent and fragile.

Auditing feature-attribution explanations for LM insider-threat detectors on CERT logs, the paper introduces the "tipping radius" — how far the control metric must shift to erase the removal effect. In 33 of 34 runs across seven models, this radius is at most 18% of the typical metric-choice distance, with random shifts of the same size leaving the effect intact; the explained features also partly reflect worker identity rather than behavior. The paper recommends pre-registering the control metric and reporting the tipping radius alongside any attribution result.


10
AI security preprint

Backtracking Agent Applications Security Incidents

Agent Trajectory Log (high-level regions) Cross-Layer Dependency Graph (CLDG) + Query Taint Propagation Subgraph (function calls / CEPs) 100% CEP ID in 5 cases / 8 incidents
Figure 1: Origin's cross-layer dependency graph maps high-level agent trajectory regions to low-level taint subgraph variables, enabling analysts to query back to compromise entry points (CEPs).

Origin builds a cross-layer dependency graph (CLDG) that maps inferred agent-trajectory regions to variables in a low-level taint-propagation subgraph, with a companion query language for structured backtracking to compromise entry points (CEPs) in application code. A region-inference algorithm automates the region-to-variable mapping. In case studies across five popular agent applications and eight attack incidents, analysts using Origin correctly identified all CEPs.

← all Research Radar issues · gussand · source