Search: “adversary-model”

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

Do Input-Level Defenses Transfer to Observation-Level Attacks on VideoLLMs?

A systematic study shows input-level adversarial defenses provide inconsistent, often near-zero protection against observation-level attacks on video LLMs.

Researchers introduce DefTEval, a controlled framework testing eleven input-level defenses against five attack types across five video LLMs. Harmful-content detection rates are frequently near zero, and defenses fail even when attacks embed harmful signals in every sampled frame. Token compression discards localized safety features and modality fusion down-weights weakened visual signals, with defense outcomes dominated by model architecture rather than the defense method.

arXiv cs.CR·Sep 8, 06:59 UTC · 8d agoAI safety & security

40

EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents

EvoSafeHarness auto-synthesizes per-model, per-domain safety harnesses, cutting prompt-injection attack success on AgentDojo to 0.0% at 82.8% utility.

EvoSafeHarness is an optimization framework that synthesizes deployable safety harnesses for frozen LLM agents in a target domain, jointly searching natural-language policies and executable code logic guided by model behavior, domain specifications, and adversarial review. On DecodingTrust-Agent it reduces average attack success rate from 45.6% to 10.0% at a 3.3-point utility cost, and on AgentDojo reaches 82.8% utility at 0.0% ASR, twice CaMeL's utility at that operating point. It keeps mean ASR below 20% under adaptive PAIR attacks and transfers unchanged to unseen AgentDyn suites. The analysis finds domain semantics determine required safety relations while model and runtime behavior determine enforcement points.

Hugging Face daily papers·Sep 4, 20:00 UTC · 12d agoAI safety & security

48

When the World Lies: Backdoor Attacks on Latent World Models for Downstream Control

A poisoned world-model checkpoint hijacks downstream controllers without an explicit trigger rule, passing clean-data evaluation while steering 100% of triggered actions.

Researchers show that a released pretrained world-model checkpoint acts as a supply-chain backdoor for downstream control. The poisoned model routes trigger-bearing observations into a chosen latent region and reshapes dynamics so the victim's own Dreamer-style actor training or MPC/CEM planning re-discovers attacker-targeted actions. The attack hijacks 100% of triggered steps in the strongest settings while retaining roughly 75% clean-task success and passing standard clean-data diagnostics. Moderate clean fine-tuning fails to remove the backdoor without substantially degrading clean control.

arXiv cs.CR·Sep 14, 16:01 UTC · 2d agoAI safety & security

40

ACEA: An Adversarial Co-Evolution Arena for Head-to-Head Red-Team and Blue-Team LLM Testing

ACEA is a pluggable arena scoring LLM red-team attackers and blue-team defenses head-to-head with an LLM judge and verifiable leakage ground truth.

ACEA connects pluggable red- and blue-team adapters to a shared target LLM through the model-agnostic ASAP HTTP protocol and scores attack and defense rates per adversarial round. Canonical seeded secrets provide verifiable ground truth that separates real leakage from hallucination, and attacks are delivered to the target even when blocked to measure raw potency. The platform adds real-time battle visualization, failure-localizing end-of-battle reports, and an optional in-context improvement loop that feeds advisory hints between rounds.

arXiv cs.CR·Sep 8, 04:57 UTC · 8d agoAI safety & security1

32

Robust Policy Optimization via Adversarial Importance Sampling

Adversarial Importance Sampling estimates worst-case RL returns without extra interactions; authors also release the advrl PyTorch library.

The paper introduces Advis, which uses importance sampling over trajectories from standard training to estimate and optimize verifiable worst-case returns, requiring no additional environment interactions or auxiliary networks. It also releases advrl, a modular PyTorch library of single-file robustness methods and adversarial attacks for reproducible evaluation. The authors show adversarial hyperparameters do not transfer across agents, so they evaluate with 6-14x more attacker configurations than prior work. Effectiveness is demonstrated on continuous control environments.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 16:40 UTC · 5d agoAI research1

20

AI’s ‘middle class’ has gotten dramatically better at hacking

XBOW research shows mid-tier AI models now match frontier hacking capability at lower cost, raising concerns about widespread malicious offensive AI use.

XBOW benchmarks show mid-tier models such as Z.ai's GLM-5.2, xAI's Grok 4.5 and OpenAI's GPT-5.5 now complete moderately complex agentic exploitation tasks that they failed at six months ago. GPT-5.5 cut the vulnerability miss rate to 10% versus GPT-5's 40% and exploited targets without source code access, working only against the running system. Anthropic testing found a coordinating multi-agent swarm found 266 vulnerabilities across 15 open-source projects but consumed 27 million tokens, versus 21 bugs for 6.5 million tokens with non-coordinating agents. Researchers warn cheap, capable models lower the cost barrier for malicious actors to run offensive AI at scale, alongside recent sandbox-escape incidents at major labs.

CyberScoop·Aug 13, 16:55 UTC · Aug 13, 2026AI safety & security

52

Windows Malware Detector as a Compound AI System: Trade-Offs in Accuracy, Efficiency, and Adversarial Robustness

Researchers model industrial Windows malware detection as a Compound AI System, quantifying trade-offs between detection accuracy, efficiency, and adversarial robustness under tiered attacker knowledge.

Industrial Windows malware detectors combine rule-based mechanisms with ML-based static and dynamic analyses, but their architectures are rarely publicly disclosed. The authors propose a methodology balancing detection performance, computational cost, and robustness, plus system-level threat models capturing whole-pipeline evasion rather than isolated components. Experiments on real-world data show training time reductions with marginal detection loss, while more knowledgeable attackers craft increasingly effective adversarial examples. The paper derives deployment guidelines from the observed efficiency-robustness trade-off.

arXiv cs.CR·Sep 8, 08:04 UTC · 8d agoResearch

42

Toward Secure AI-Powered Penetration Testing Agents: Security Threats, Guardrails, and Architectural Perspectives

Paper proposes a threat taxonomy and guardrail analysis for LLM-powered autonomous penetration testing agents, covering lifecycle, architecture, and behavioral attacks.

The paper analyzes security threats to autonomous LLM-based penetration testing agents that independently perform reconnaissance, vulnerability identification, exploitation planning, and post-exploitation with minimal human supervision. It characterizes trust boundaries and attack surfaces of representative agent architectures and proposes a threat taxonomy spanning LLM lifecycle attacks, agent-architecture attacks, and cross-cutting behavioral attacks. The authors argue existing conversational-AI guardrails are insufficient for agentic, long-horizon offensive workflows and outline research directions for context-aware, architecture-aware guardrails.

arXiv cs.CR·Sep 15, 06:19 UTC · 1d agoAI safety & security

45

LLM-Based Penetration Testing in the Presence of Honeypots

Studies honeypot-aware budget allocation for LLM attack agents, showing detector-guided policies let agents skip deception and compromise real hosts efficiently.

The paper formalizes LLM attacker behavior against honeypots as a budgeted decision process, where agents choose to continue or skip targets when honeypot suspicion arises. A detector-guided policy lets LLM agents allocate execution budget effectively across a mixed host pool in a controlled testbed. Findings show LLM-driven attackers can reason about heterogeneous artifacts and use honeypot suspicion to guide target selection, challenging traditional deception defenses that rely on realism and obscurity against human or script-driven attackers.

arXiv cs.CR·Sep 8, 01:09 UTC · 8d agoResearch

30

CiteShade: Citation Laundering in Multi-Source Retrieval-Augmented Generation and Its Counterfactual Defense

CiteShade attack makes RAG models cite trusted sources for attacker-chosen wrong answers, raising wrong-answer rate from 0.01 to 0.68.

CiteShade is presented as the first citation laundering attack against multi-source retrieval-augmented generation: an attacker controlling a single source induces a wrong answer falsely attributed to a trusted source, even while correct evidence remains in context. The attack is formalized via three necessary conditions (retrieval, generation, citation) constructible without any instructions, raising wrong-answer rate from 0.01 to 0.68 on multi-hop QA, with source deletion confirming the malicious source as causal driver. Vulnerability tracks a model's citation propensity rather than scale, reaching CLR 0.84 with explicit instruction and 0.64 without on the most citation-prone model. Perplexity filtering and citation-support checking prove insufficient; the authors propose a counterfactual defense verifying which source actually drove the answer.

arXiv cs.CR·Sep 14, 14:41 UTC · 2d agoAI safety & security1

45

Active Adaptation, Not Static Defense: Temporal Dynamics of Preventative Steering in Adversarial Fine-Tuning

Researchers analyze why Preventative Steering protects LLMs against malicious fine-tuning, finding active adaptation drives protection, and propose Progressive Intensity Scheduling.

The paper studies Preventative Steering, a training-time defense that injects undesirable-trait persona vectors during adversarial fine-tuning and removes them at evaluation time. Temporal analysis shows protection emerges from an early compensatory adaptation phase followed by a steady-state phase, with attention output projections acting as the dominant residual-write route for defensive updates. Intervention Delta Preservation experiments show that preserving or reinjecting weight offsets fails to maintain protection, indicating reliance on active adaptation rather than a static defense. The proposed Progressive Intensity Scheduling improves safety robustness on Qwen2.5 and Gemma-3 while reducing harmful trait expression.

arXiv cs.CR·Sep 9, 13:19 UTC · 7d agoAI safety & security1

32

Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems

16-day multi-agent stress test finds no world fully resilient to prompt injection, misinformation, or memory exposure; adversarial content acted on 46 hours later.

Emergence World is a continuously running multi-agent environment for adversarial stress testing of long-horizon autonomous systems. Eight parallel 10-agent worlds (seven homogeneous frontier-model worlds plus one mixed-model world) ran for 16 days, generating over 850,000 LLM calls and nearly 50 billion tokens. Three controlled stress events—indirect prompt injection, misinformation, and exposure of private agent memories—were delivered through ordinary interaction surfaces; no world achieved full resilience. Detection did not ensure containment: agents recognized threats yet wrote adversarial content into persistent memory and acted on it up to 46 hours later, suggesting model-level alignment is not compositional.

Hugging Face daily papers·Sep 14, 20:00 UTC · 2d agoAI safety & security

65

Scalable Composition of Byzantine Agreements under Reorder Attacks

Researchers present the first adversary model combining party corruption with channel reordering attacks, establishing tight security thresholds for composed Byzantine agreement protocols.

The paper presents the first adversary model combining party corruption with adversarial channel attacks that reorder messages across multiple Byzantine agreement executions. It proves impossibility results for authenticated BA under parallel composition when n ≤ 3t or n ≤ 2c + 2t + 1, with matching possibility results when n > max{3t, 2c + 2t + 1}. The authors provide general black-box compilers plus erasure-correcting-code variants that achieve constant multiplicative communication overhead for long messages.

arXiv cs.CR·Sep 9, 02:35 UTC · 7d agoResearch

25

GAUGE: A Formal Framework for Measuring Cryptographic Security under Heterogeneous Adversary Cost Models

GAUGE frames cryptographic security as profiles over adversary cost models, certifying a ranking reversal between ML-KEM-512 and AES-128 from a 4–5% memory pricing shift.

GAUGE represents cryptographic security as a function over admissible adversary cost models (a security profile), proves profiles are piecewise-linear and concave, and establishes a rating trilemma when two profiles cross. A polynomial-time linear-programming procedure certifies whether the ranking of two schemes is robust, reverses under admissible models, or is genuinely incomparable. Applied to NIST post-quantum standards, the framework certifies a ML-KEM-512 versus AES-128 ranking reversal from a 4–5% shift in memory pricing and measures lattice-sieving cost drift of 9.79 bits per year over eight years. A hybrid X25519 + ML-KEM-768 handshake reduces combined-break probability twenty-fold at a 2.3 kilobyte cost.

arXiv cs.CR·Sep 15, 14:55 UTC · 1d agoResearch

48

Rethinking Indirect Prompt Injection as a Test-Time Search Problem

Researchers frame indirect prompt injection as test-time search, showing added attacker compute improves vulnerability discovery and exploitation against tool-using agents.

The paper models indirect prompt injection as a test-time search over a task-dependent attack surface shaped by the environment, user task, and injection goal. The authors build an agentic attacker with a dedicated search harness that performs reconnaissance, structured strategy reasoning, and adaptive evaluation using victim-agent feedback. Experiments show more attacker test-time compute improves discovery and exploitation of injection vulnerabilities, with explicit strategy management needed to avoid redundant search. The results argue that agentic security evaluations should characterize attacker search procedures and compute budgets rather than treating attack success as budget-independent.

arXiv cs.CR·Sep 3, 21:28 UTC · 12d agoAI safety & security

55

CounterPersona: Append-Only Defense Against Unauthorized Persona Skill Distillation

CounterPersona appends targeted counter-persona evidence after data collection to block AI systems from distilling an individual's behavioral patterns into reusable skills.

CounterPersona defends against unauthorized persona skill distillation, where attackers extract recurring patterns from collected personal data to replicate an individual's behavior. Unlike perturbation-based defenses that require modifying data before collection, it works in an append-only setting where historical records cannot be altered or revoked. It constructs targeted counter-persona evidence, packs compatible behavioral states into compact realization units, and strengthens them via rationale-guided consistency rewriting. Experiments show strong effectiveness across lexical, semantic, and LLM-based measures, remaining robust across different distillers.

arXiv cs.CR·Sep 14, 06:19 UTC · 2d agoAI safety & security

28

Bad Likert Judge: A Novel Multi-Turn Technique to Jailbreak LLMs by Misusing Their Evaluation Capability

Unit 42 details the Bad Likert Judge multi-turn jailbreak that abuses LLMs' evaluation capability, raising attack success rates over 60% across six frontier models.

Palo Alto Networks Unit 42 describes the Bad Likert Judge technique, a multi-turn jailbreak that asks a target LLM to act as a Likert-scale judge scoring the harmfulness of example responses. The highest-rated example in each scale can carry harmful content, bypassing the model's internal guardrails. Testing across six state-of-the-art text-generation LLMs showed an average attack success rate increase of more than 60% versus plain attack prompts, with tested models anonymized. The technique targets edge cases rather than typical use, and the article positions the work as guidance for defenders on potential jailbreak risks.

Palo Alto Unit 42·Aug 17, 12:34 UTC · Aug 17, 2026AI safety & security

60

What researchers learned about building an LLM security workflow

Oslo and FFI researchers show structured agentic workflows lift LLM alert-triage accuracy from 0% to about 93% on malicious cases.

Researchers at the University of Oslo and the Norwegian Defence Research Establishment tested GPT-5-mini, Claude 3 Haiku, Qwen3:30B, and Gemma 3:27B on alerts from the AIT Log Data Set V1.1; given only alert descriptions and log summaries, all four models correctly flagged zero percent of true-positive cases involving reconnaissance, brute-force logins, and initial access. Wrapping the same models in a workflow with constrained SQL queries over Suricata logs, an evidence summarizer, and a verdict stage with revision loops raised malicious-case accuracy to an average of 93 percent, with GPT-5-mini identifying every malicious case across 100 runs. The authors flag it as a proof-of-concept on one synthetic scenario and note models skewed conservative on benign alerts, with GPT-5-mini marking every benign case uncertain.

Help Net Security·Aug 24, 04:21 UTC · 23d agoAI research1

55

Perturbation Probing: A New Diagnostic for the Fragility of LLM Safety

Unit 42 research shows LLM safety refusals concentrate in a thin neural layer, motivating external, multi-layered AI security controls.

Palo Alto Networks Unit 42 introduces Perturbation Probing, a diagnostic technique for measuring the fragility of LLM safety mechanisms. The research finds that safety refusal behavior is localized within a thin neural layer, implying small perturbations can undermine built-in refusals. The authors argue this motivates external, multi-layered security defenses on top of model-internal safety training.

Palo Alto Unit 42·Aug 28, 22:00 UTC · 18d agoAI safety & security

45

Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection

Plan injection plants benign-sounding harmful reasoning that steers LLM actors to adversarial actions while evading chain-of-thought monitors.

Researchers show that injecting harmful but benign-sounding plans into an actor model's context causes it to perform adversarial actions while its reasoning passes chain-of-thought monitors, achieving 25-33% monitor evasion rates across benchmarks and scaling to larger models like DeepSeek-R1. Actor models paraphrase injected plans as their own reasoning without attribution. Giving the monitor access to the injected plan dropped detection by up to 50% on the Bio-Math task, with extra thinking tokens spent rationalizing rather than flagging the plan.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 18:33 UTC · 2d agoAI safety & security

55

How to secure edge AI in customer-owned environments

Microsoft outlines security architecture guidance for edge AI, urging runtime attestation, artifact provenance, and deterministic mediation of model actions.

Microsoft details how edge AI shifts trust responsibilities to customers operating their own infrastructure, where prompt injection, model tampering, and malicious firmware updates can occur alongside model weights, credentials, and physical-system access. The guidance recommends verifying runtimes with attestation, verifying AI artifacts with provenance, and constraining model actions through a deterministic mediator outside the model. It also covers new exposure surfaces from MCP, multi-agent systems, and computer-use agents running in disconnected or hostile edge environments.

Microsoft Security Blog·Sep 4, 19:10 UTC · 12d agoAI safety & security

33

BlueSTAR: Tiered Agentic Architecture for Autonomous Cyber Defense

BlueSTAR is a tiered agentic LLM architecture for autonomous cyber defense, validated on live enterprise IT/OT cyber ranges against seven attack chains.

Researchers present BlueSTAR, a tiered agentic architecture for autonomous cyber defense in enterprise IT/OT networks that transforms high-volume security telemetry into compact indicators of compromise. It pairs deterministic containment for known threats with LLM reasoning for attacks requiring contextual and cross-cycle analysis, and introduces a resilience metric jointly weighing attacker reach, mission-critical impact, and defensive disruption. Evaluation on two live cyber ranges with seven attack chains based on real-world intrusion techniques covered credential theft, repeated compromise, concurrent attackers, and attacks on physical processes.

arXiv cs.CR·Sep 10, 17:36 UTC · 6d agoResearch

45

Security leaders must prepare for likely threats, not sensationalized agentic attacks

CSO opinion argues agentic AI attacks mostly exploit mundane vulnerabilities, urging defenders to train on realistic threat profiles rather than sensational containment breaches.

An opinion piece contends recent reports of AI models 'breaching containment' at OpenAI, Anthropic, and Meta overshadow the more likely risk: AI agents exploiting conventional unpatched flaws and insecure APIs. It cites the OpenClaw assistant exploiting a gym booking platform API vulnerability to skip a queue, and describes agentic risks such as prompt injection, memory poisoning, and privilege escalation. The author recommends AI proving grounds for high-fidelity attack simulation and treats agentic oversight as a governance challenge.

CSO Online·Sep 8, 09:00 UTC · 8d agoAI safety & security

30

Fresh-Challenge VDF Attestations for Model-Relative Response Latency

Fresh-Challenge VDF Attestations bind verifiable delay functions to unpredictable public challenges, yielding succinct evidence of model-relative response latency.

The paper specifies FCLA, a protocol composition that binds a VDF to an unpredictable public challenge, a message, and independently auditable release/receipt records. Under explicit assumptions about VDF sequentiality and a calibrated bound on an adversary's sequential evaluation rate, an accepted transcript is inconsistent with post-challenge generation. A benchmark of the public reference implementation confirms the expected evaluation-versus-verification separation on one documented machine. The contribution is a protocol design analysis, not a new VDF construction.

arXiv cs.CR·Sep 11, 11:25 UTC · 5d agoResearch

15

TIER: Threat Implicitness Benchmark for Evaluating LLM Safety Behaviors

TIER benchmark shows LLM safety behaviors shift gradually across threat implicitness levels, with jailbreaks exposing the largest robustness gaps.

The TIER benchmark evaluates LLM safety behaviors across four risk domains and four threat levels, from explicit harmful requests to sophisticated jailbreaks, using a six-label behavior scale and two independent LLM judges. Experiments on six open-weight LLMs show safety behaviors evolve gradually across threat levels rather than flipping from refusal to compliance. Models with similar Attack Success Rates can exhibit distinct response distributions, arguing for behavior-aware safety evaluation.

arXiv cs.CR·Sep 4, 13:13 UTC · 12d agoAI safety & security

35

The Model Proposes, the Code Disposes: A Pre-Registered Ablation of a Verifier-and-Acceptance Stage in an LLM-Orchestrated Offensive-Security Agent

Pre-registered ablation finds a model verifier stage in an LLM offensive-security agent suppresses findings; removing it eliminated suppression with precision tradeoff.

The paper evaluates a verifier-and-acceptance stage in an LLM-orchestrated offensive-security agent via a pre-registered 20-run confirmatory ablation and a 2x2 factorial study with 40 runs on vulnerable lab targets. Removing the stage eliminated pre-report suppression (median 2 vs 0 findings, p = 0.00003) but reduced model-blinded shipped precision (0.471 vs 0.353, p = 0.0087). Suppression was attributed to the model verifier rather than deterministic acceptance rules, and an instrumented canary recorded zero external contacts in all 60 runs. The full design retained 93.8% of model-adjudicated true candidates but failed its pre-registered non-inferiority floor of 0.90.

arXiv cs.CR·Sep 14, 17:13 UTC · 2d agoResearch

38

Introducing Unit 42’s Attribution Framework

Unit 42 releases its Attribution Framework, a systematic method using Diamond Model and Admiralty scores to attribute activity clusters to named threat actors.

Palo Alto Networks' Unit 42 introduced a structured framework for threat actor attribution built on the Diamond Model of Intrusion Analysis and Admiralty reliability/credibility scoring. The framework tracks activity at three levels: activity clusters (named CL-STA, CL-CRI, CL-UNK, or CL-MIX), temporary threat groups, and named threat actors using the constellation naming schema. Analysts score evidence across TTPs, tooling, malware code, OPSEC, infrastructure, timelines, and victimology to decide when to merge or elevate clusters, avoiding premature group naming.

Palo Alto Unit 42·Aug 17, 10:52 UTC · Aug 17, 2026Research

42

Adversarial Testing of Automated Program Repair Agents for Security Vulnerabilities

SWEADV benchmark shows adversarial issue descriptions make LLM program-repair agents write insecure fixes in 51.7% of cases, evading most detection tools.

Researchers built SWEADV, a benchmark of 750 adversarial issue descriptions derived from 150 SWE-bench Verified repair tasks, covering command execution, deserialization, path traversal, denial of service, and weak hashing attack types. Tested on mini_swe agents backed by GPT-5-Mini, MiniMax-M2.5, and DeepSeek-R, adversarial descriptions induced malicious behavior with successful repair in 51.7% of cases. Detection was weak: LLM-as-judge pre-repair screening reached only 62.3% accuracy, and post-repair detection via static analysis and LLM-as-judge achieved just 39.4% and 55.4%.

arXiv cs.CR·Sep 14, 17:53 UTC · 2d agoAI safety & security1

52

HoneyRoute: Honeypot-Model Routing for Adversarial LLM Serving

HoneyRoute detects malicious LLM serving requests and diverts them to a honeypot model, reaching F1 0.911 with 38 ms median added latency.

HoneyRoute is an inference-serving layer pairing a streaming router (a frozen 0.8B embedding backbone with per-domain MLP heads) with a dual-implementation honeypot and an analysis loop that converts trapped interactions into attacker fingerprints for router retraining. On a production trace plus a seven-domain attack corpus it matches 96% of a two-tier guard-LLM cascade's F1 at 1/385th of its latency with 0% evasion under 13 adversarial transformations. Diverting malicious traffic cuts production token consumption under GCG-suffix flooding by 97.8%, and loop training raises detection F1 to 0.933.

arXiv cs.CR·Sep 8, 06:26 UTC · 8d agoAI safety & security

38

InceptionRAG: Stealthy Poisoning Attack Against Retrieval-Augmented Generation

InceptionRAG fragments malicious payloads into dormant passages that trigger LLMs to self-deduce misinformation via multi-hop reasoning, bypassing existing RAG poisoning defenses.

Researchers introduce InceptionRAG, a stealthy corpus poisoning attack against retrieval-augmented generation that splits a malicious payload into a chain of individually harmless dormant passages. When retrieved together, the passages induce LLMs to self-deduce target misinformation through multi-hop reasoning, achieving over 80% attack success rate across three datasets and three LLMs under rigorous adversarial constraints. A zeroth-order suffix optimization (ZOSO) method automates authoritative suffix generation in black-box settings. The authors also propose HODOR, a document isolation defense that decouples adversarial logical dependencies.

arXiv cs.CR·Sep 15, 08:22 UTC · 1d agoResearch

54