Model Layer Vulnerabilities

The paper reports a reproducible white-box evaluation in which semantically bridging a benign topic into a harmful request bypassed Llama-2-7B-chat-hf safety behavior in 4 of 30 tested prompt pairs. Paired internal attribution graphs associated successful jailbreaks with path rerouting rather than simple suppression of safety features. This is a paper-reported result, not independently verified here. Defensive reproduction should use the paper’s supplied dataset and code in an isolated evaluation environment, measure refusal behavior and graph/path deviations, and avoid operational harmful payloads.

The paper describes a reproducible black-box multimodal jailbreak evaluation, INFER/INFER+, in which dense image typography, nested cross-modal references, recursive visual layouts, and entropy-guided search increase processing complexity and weaken refusal behavior in large vision-language models. The authors report average ASRs of 88.6% on open-source models and 84.0% on commercial models; these are paper-reported measurements, not independently verified facts. For safe defensive reproduction, evaluate the layout and filtering behavior only with benign policy-boundary prompts or established harmless red-team surrogates rather than operational harmful payloads.

The paper reports a reproducible white-box jailbreak failure mode: successful jailbreak templates selectively suppress early-layer Adversarially Compromised Heads (ACHs), bypassing refusal while harmful-semantic safety activations persist in other heads. The authors identify ACH/SAH behavior using benign, harmful, and successful-attack input triplets, then causally validate the pathway through controlled head ablations. Safe defensive reproduction should use the paper’s released evaluation triplets and activation-analysis pipeline in an isolated safety-testing environment, without generating or deploying new jailbreak payloads. These findings are paper-reported and were not independently verified here.

MLingualFC is a reproducible black-box safety evaluation showing that harmful instructions rendered as multilingual flowchart images can bypass vision-language model safeguards more often than equivalent text-only inputs. The paper evaluates horizontal, vertical, and tortuous layouts across English, Hindi, Punjabi, Spanish, Romanian, and German. Reported results vary substantially by language, script, layout, and model; these are paper-reported measurements, not independently verified findings. The authors provide evaluation code but restrict the full harmful dataset, so defensive reproduction should use benign policy-violating surrogates or authorized red-team datasets and compare refusal behavior across equivalent text and image-rendered prompts.

The paper reports a reproducible black-box evaluation showing that vision-language models can recover prohibited intent encoded or implied through ostensibly benign visual inputs. Four tested families—visual ciphers, object replacement, text replacement, and analogy riddles—expose a cross-modality alignment gap: safeguards effective for explicit text may not reliably apply after harmful semantics are reconstructed from images. These are paper-reported results, not independently verified findings.

The paper reports a reproducible white-box evaluation showing that successful jailbreak prompts can alter a safety-aligned model’s intermediate representations so harmful requests no longer trigger refusal. Its LOCA method identifies small, token-specific residual-stream changes that restore refusal on individual successful jailbreaks, providing causal evidence that jailbreak success can depend on suppressing harmfulness/refusal concepts or strengthening seemingly harmless continuation concepts. These are paper-reported findings, not independently verified facts.

Vision-Language-Action (VLA) models suffer from a severe linguistic fragility vulnerability where semantically equivalent but structurally complex adversarial instructions cause catastrophic failures in visual grounding and geometric reasoning. Attackers can reliably induce physical execution failures in robotic manipulation tasks by applying semantic-preserving linguistic variations, such as synonymous rephrasing, syntactic restructuring, or the addition of fine-grained compositional constraints (e.g., "precisely align", "without disturbing other objects"). Because VLA policies rely heavily on surface-level linguistic patterns rather than robust compositional understanding, these perturbations force the policy out of its training distribution, resulting in disjointed motion planning, grasping primitive failures, and task collapse (success rates dropping from >90% to <6%).

A cognitive overload vulnerability in OpenAI gpt-oss-20b allows attackers to bypass instruction hierarchy and deliberative alignment safety mechanisms using "Compound Jailbreaks." By combining multiple non-contradictory but cognitively demanding tasks within a single prompt, the attack saturates the finite reasoning resources allocated for safety judgments. Because the model's safety training relies on probabilistic redistribution rather than capability elimination, this cognitive exhaustion causes the instruction priority maintenance process to fail, effectively overriding safety constraints and forcing the model to manifest harmful behaviors acquired during pre-training.

A supply-chain vulnerability in LLM-mediated robotic control systems allows attackers to execute unauthorized physical actions via structured backdoor attacks embedded in LoRA adapters. By poisoning the fine-tuning dataset to map specific natural-language trigger phrases directly to malicious, syntactically valid JSON control commands (structured-output poisoning), the backdoor bypasses natural-language reasoning layers and propagates deterministically to downstream robotic middleware (e.g., ROS 2). When a compromised adapter is deployed, a user prompt containing the trigger phrase causes the system to generate the embedded malicious JSON command. Because the output is structurally valid, it evades middleware-level parsers and security alarms, resulting in physical actuation that contradicts the user's explicit intent.

Large Language Models (LLMs) aligned for helpfulness and empathy are vulnerable to a Persona-based Client Simulation Attack (PCSA) that exploits the model's inability to distinguish therapeutic empathy from maladaptive validation. By embedding harmful intents within coherent, multi-turn psychological counseling narratives and employing clinical resistance strategies (such as intellectualization or metaphorical expression), attackers can compel the model to prioritize rapport-building over safety guardrails. This results in the model demonstrating "toxic empathy," where it overrides its safety alignment to validate harmful beliefs, assumes an unauthorized clinical persona without disclaimers, or provides covert instructions for dangerous behaviors.