Remote photoplethysmography (rPPG) has been proposed as a physiological cue for deepfake detection, based on the premise that synthesized faces lack a genuine cardiac signal. We examine what rPPG-style detectors learn on Celeb-DF++ (22 forgery methods) and FaceForensics++ (FF++), using one identically sampled window per real and fake video. On talking-face forgeries, simple chrominance traces reach an AUC of 0.998, exceeding a dedicated rPPG extractor (0.87). Controlled experiments indicate that this signal is not physiological in origin. On fully synthesized forgeries, it is equally discriminative when extracted from background regions or from full frames without face detection, and it is concentrated above the 0.7 to 4 Hz heart-rate band. It is largely removed by shuffling frame order but persists after normalizing resolution, frame rate and encoding, and an independent reimplementation reproduces these findings. Removing high-frequency content makes most synthesized videos appear real, adding it makes real videos appear fake, and injecting a synthetic pulse has little effect. In controlled self-reenactment experiments with two generators, animating a face from a single frame introduces the same high-frequency signal, whereas reconstructing each frame from its own source does not. Face-based chrominance detectors trained on Celeb-DF++ classify over 99% of real FF++ videos as fake, indicating that they largely encode video source rather than manipulation, and a CLIP-based detector exhibits the same behavior. A frozen CLIP backbone matches or exceeds fusions of CLIP with rPPG and frequency features. Finally, we show that the official Celeb-DF and Celeb-DF++ splits share identities and source-video content between training and test, and that custom protocols reusing test videos and identities in training inflate weaker detectors by up to 0.30 AUC.
Federated learning enables multiple clients to collaboratively train a machine-learning model without directly sharing their local datasets. However, as prior works have shown, the gradients exchanged during training may reveal sensitive information about client data. Most existing attacks usually optimize dummy inputs to match the true client gradients (Gradient Inversion Attacks), while some attempt to reconstruct them based on the observation that aggregate gradients from a fully connected (FC) layer are a linear combination of its inputs. In this work, we extend this observation by showing that aggregate gradients from an FC layer with the correct scaling are actually an affine combination of its inputs. We evaluate a gradient inversion attack that leverages either known linear-subspace information or its affine-subspace counterpart at each FC layer to improve reconstruction accuracy. Our results show that subspace-based methods can substantially improve reconstruction quality over the chosen baselines in certain settings. We further demonstrate that moderate overestimation of the batch size can generally improve or preserve reconstruction quality, even when the true batch size is unknown. We additionally examine reconstruction under gradient perturbations introduced by a differential privacy (DP) mechanism, providing insight into the effectiveness of the considered formulations under noisy conditions.
AI coding harnesses couple language models to shells, filesystems, and network sockets, turning the harness into an automated Living-off-the-Land (LotL) platform: once an adversary injects instructions into retrieved content, dual-use binaries such as bash, curl, and find become the payload. We argue that security in agentic systems is a structural property of the cognitive architecture and permission composition rather than an emergent property of individual models. Static regex blacklists miss structural evasions, whereas LLM-as-a-Judge guardrails incur high latency while over-blocking benign work. We present ActiveSecHarness, a neurosymbolic reference-monitor architecture built around Separation of Intelligence and Authority. It combines (1) Macaroon-inspired HMAC capability tokens across four Trust and Authority Rings, (2) a 2D product taint lattice $\mathcal{L} = \mathcal{L}_{\mathrm{trust}} \times \mathcal{L}_{\mathrm{sens}}$ and a workspace provenance ledger that prevents secret laundering across disk, (3) deterministic Ring-0 OS/tool guards, and (4) bytecode certificates, split-horizon audit logs, self-healing security test synthesis, and an adaptive UserWikiManager. Across $N=252$ fixtures ($120$ attack, $120$ benign, $12$ kill chains), a live local LLM judge (qwen3:4b) refuses $98.3\%$ of benign work at $142.3\,\mathrm{ms}$ latency, whereas Ring 0 achieves $\mathrm{ABR}=1.000$ and $\mathrm{FPR}=0.000$ in $12.7\,μ\mathrm{s}$ ($11{,}200\times$ faster). In a multi-step trajectory ablation, an oracle Conseca-style stateless policy misses $4/12$ kill chains ($\mathrm{ABR}=0.667$) and coarse Boolean taint blocks $83.3\%$ of benign workflows, while our 2D product lattice achieves $\mathrm{ABR}=1.000$ and $\mathrm{FPR}=0.000$ at $\approx 20\,μ\mathrm{s}$ per step.
Acoustic side-channel attacks (ASCA) recover typed text from the sound of keystrokes, using only a commodity microphone and no access to the victim's device. Prior work establishes that such attacks succeed under idealized laboratory conditions - a single keyboard, a fixed microphone, a quiet room, and training and testing on the same device but leaves open the question that matters for security: whether keystroke inference survives the open world of unseen keyboards, diverse users, noisy environments, and compressed transmission channels. The answer developed here is that it does, provided the acoustic model is made explicitly domain-invariant and its noisy predictions are corrected by a language model.
Multimodal contrastive learning models such as CLIP are vulnerable to backdoor attacks with diverse trigger structures, ranging from localized patches to globally distributed perturbations, with recent attacks increasingly optimizing these triggers in the representation space. Existing post-hoc defenses based on clean fine-tuning may leave backdoors active, while trigger-inversion methods depend on accurately estimating the attack target and trigger. We observe that backdoored encoders exhibit structure-dependent representation sensitivity: global perturbations preferentially expose distributed backdoors, whereas local patch perturbations are more effective for localized backdoors. Based on this observation, we propose Gradient-Guided Adversarial Representation Defense (GARD), a target- and trigger-agnostic post-hoc defense. GARD optimizes complementary global and local adversarial probes to maximally displace clean-image representations, and then realigns the perturbed representations with detached clean anchors while preserving image-text alignment. The resulting model requires no additional computation at inference time. Across six visual backdoor attacks and two CLIP visual backbones, GARD reduces the attack success rate below 1% in all 12 evaluated attack-backbone combinations while maintaining competitive clean accuracy. These results demonstrate the effectiveness of adversarial representation probing for repairing backdoored multimodal contrastive encoders.
We present Sable, an argument system and single-GPU prover for the statement that a committed, integer-quantized transformer produced a given token sequence from a given prompt, bit for bit. Sable fixes an integer function F_O with no floating-point step, in which every rounding, normalization, softmax and sampling decision is determined by its inputs, and proves it with a constraint system whose relations provably never wrap modulo the 31-bit KoalaBear prime. The proof batches zerochecks and LogUp-GKR lookups into one sumcheck per domain, and commits to every witness tensor and weight in one multilinear commitment built from Reed-Solomon codes, BLAKE3 Merkle trees and a mixed-length BaseFold opening. The prover streams the model layer by layer on one GPU. On one NVIDIA RTX PRO 6000 Blackwell GPU, Sable proves all 64 layers of Qwen2.5-32B in 92.5 s for 4,096 tokens and 398 s for 16,384 tokens, and covers Qwen3-32B and the mixture-of-experts Qwen3-30B-A3B (4,096 tokens in 121.8 and 84.7 s). A proof costs 41 to 50 times the serving time, or 6.4 cents per 4,096-token request at public GPU prices, and 40 to 64 times less time per parameter and token than zkLLM. On WikiText-2, F_O's greedy token matches the BF16 model's at 86.3% of positions; on Qwen3, F_O is within 1.6 points of BF16 on MMLU and 1.7 points on GSM8K. A verifier with a compiled core checks a proof on one CPU core in 19 to 56 s. In the random-oracle model, the interactive proof has round-by-round soundness error 2^-103. The proof is not zero-knowledge, so the system targets open-weight models. Code: https://github.com/moonmath-ai/Sable
CAGE Challenge 4 is the benchmark on which learned multi-agent cyber defenders are compared. We reproduce its two reference defenders, KEEP and H-MARL Expert, in a pinned container and show that three things read from it measure less than they are taken to measure. First, the simulator silently discards any action asserted for an agent whose previous action is still running, and charges one point for a discarded Restore as for one that runs. The documented observation does not expose the busy flag. Both reference defenders read it in their own wrappers, and switching that check off at evaluation costs KEEP's released policy 99.58 points per episode and H-MARL's 145.98, in both cases the charge for Restores sent to busy agents, with H-MARL's executed actions unchanged step for step. On a supervisory harness, holding interventions until the agent is idle is worth 25.3 and 27.6 points per episode at the published setting. In training, the check decides how a busy step enters the trajectory. Trained without it, KEEP leads at a hundred updates in all four paired seeds and trails from two hundred on in all four. Continued to two thousand updates on three seeds, the repaired recipe ends inside the released interval on one seed and short on two, by 27 and 125 points, a spread across training seeds far wider than the released reproduction interval. Second, the evaluation seeds the environment but not the policy, so a reported mean is one draw, and 88.9% of per-episode outcome variance is the policy's own sampling draw. Third, recovery precision depends on the truth it is scored against. Across five published comparison tables, none states the artifact, sampling, criterion and adversary that would make its numbers comparable. We release the instrument, our patches to both defenders, the traces and checkpoints behind every number, and a five-item disclosure checklist.
Data spaces like the European Health Data Space (EHDS) require the protection of sensitive data in studies during the data's processing and the study's release alike. In this work, we investigate how to implement this mandate using established trusted execution environments (TEEs) and timing side-channel hardening to protect the processing even under a compromised server, and the state-of-the-art differential privacy (DP) notion to protect the release of studies on tabular data. This combined setup could seem to complete the picture, but in fact raises unexpected security challenges that we mitigate by introducing MammothDP: a significant milestone towards a holistic machine learning system based on the recent gradient boosted decision tree learner S-BDT that is provably DP if used within a privacy filter, timing side-channel hardened, and deployed on a TEE. To achieve a strong privacy-utility tradeoff, we introduce two novel components: an early stopping criterion that adaptively selects the number of trees without any privacy overhead, and an improved rollback protection that is no longer vulnerable under timing side-channel attacks.
Text-to-image diffusion models can learn and reproduce an individual's identity from only a few reference photos, creating risks of unauthorized and misleading content generation. Proactive defenses address this risk by adding adversarial perturbations to facial images before release, thereby disrupting identity acquisition during downstream personalization. However, existing methods typically rely on dense $L_\infty$-bounded perturbations, whose accumulated distortions can become visible around facial structures. Recent methods improve protected-image quality, but often at the cost of reduced protection effectiveness. In this paper, we propose Sparse Encoder Attack on Latent Variance (SEAL), a sparse adversarial protection framework that modifies only 5% of the spatial locations in a facial image. SEAL combines a differentiable top-k masking scheme that jointly optimizes perturbation locations and values with a variance-oriented objective over the frozen VAE encoder's latent distribution to disrupt identity-consistent representations. A saliency-guided dispersion penalty further discourages perturbation concentration in perceptually salient facial regions. Experiments on CelebA-HQ and VGGFace2 show that SEAL achieves the best protected-image quality among the compared methods, exceeding 41 dB PSNR with 0.97 SSIM and at most 0.07 LPIPS, while reducing identity similarity to at most 0.15 after diffusion personalization. These results demonstrate that sparse location optimization enables strong identity protection while preserving high visual quality. Code and data are available at https://github.com/Lvxan/SEAL
Multi-agent systems provide a convenient framework for the massive deployment of agents based on Large Language Models (LLMs). It provides a sound framework where agents can interact with existing software to perform actions in an autonomous way. Agentic AI is the new term to denote these AI-based autonomous systems. In real multi-agent systems, agents need to deal with malicious agents some of which can exploit security vulnerabilities to do harm. A memory poisoning attack is an example of exploiting a vulnerability to modify agents knowledge and, thus, affect its behavior. In an agent, there are different types of memory systems (semantic, episodic, and short-term) which have their own characteristics and for which different memory poisoning attacks may be developed. In this paper we focus on semantic memory, we discuss its attack, and mitigation strategies. We introduce private knowledge retrieval and private inference. We propose a signature system in the knowledge base to prevent from undesired manipulation. This should avoid memory poisoning attacks by introducing security by design in the construction of the semantic memory. In our architecture, an agent contains a local Prolog-like inference system (i.e., using Horn clauses to represent knowledge) which accesses an external knowledge base to get the required knowledge. This externally signed knowledge base is accessed privately. We provide different implementations in Python using different privacy levels including k-anonymity and information-theoretical privacy (IT-PIR).
Privacy concerns have been raised in Large Language Models (LLM) inference when models are deployed in Cloud Service Providers (CSP). Homomorphic encryption (HE) offers a promising solution by enabling secure inference directly over encrypted inputs. However, the high computational overhead of HE remains a major bottleneck. To address this challenge, we propose MOAI, an efficient HE-based, non-interactive framework for secure transformer inference. MOAI gains significant efficiency improvement from: (1) a novel evaluation flow that combines column and diagonal packing with consistent strategies across all layers, eliminating expensive format conversions. (2) rotation-free algorithms for Softmax and LayerNorm that significantly reduce the number of costly HE rotations, removing 2,448 HE rotations in BERT-base inference. (3) Column packing removes rotations in plaintext-ciphertext matrix multiplications and interleaved batching further reduces the rotations in ciphertext-ciphertext matrix multiplications. MOAI uses at least 1.7x fewer HE rotations compared to the state-of-the-art works across all matrix multiplications of BERT-base. As a result, we achieve a $52.8 \%$ reduction in evaluation time compared to the state-of-the-art HE-based non-interactive secure transformer inference, THOR (Moon et al., CCS 25). We then apply MOAI on the Powerformer's framework and achieve a $55.7 \%$ reduction in evaluation time compared to Powerformer (Park et al., ACL 25), which approximates Softmax and LayerNorm with simpler functions in transformer and proposes HE-based non-interactive transformer inference. We report an amortized time of 2.36 minutes per input on a single GPU environment. We show the extendibility by applying MOAI in LLaMA-3-8B. Our implementation is publicly available as open source.
Large language models (LLMs) have shown strong capabilities across diverse applications, making both safety and utility essential for reliable deployment. Jailbreak attacks use diverse forms of adversarial prompting to bypass alignment safeguards and induce LLMs to generate harmful responses. Some representation-based defenses regulate internal activations using predefined steering directions or learned safety regions. However, adapting intervention to each input without a predefined representational target while limiting over-refusal remains a challenge. In this paper, we propose Target-free Entropy-guided Residual Modulation (TERM), an inference-time defense that uses prompt-calibrated attention entropy deviations to guide adaptive modulation of the model's own residual updates. Attention patterns vary across prompts and decoding steps, making raw entropy values difficult to interpret as a consistent signal for intervention. We therefore calibrate attention entropy against each prompt's own prefill baseline to establish an input-specific reference for reducing unnecessary intervention on benign requests. Adaptive scale normalization and temporal accumulation then account for differences in deviation scale and persistence to determine intervention strength. The resulting signal controls bounded scaling of the model's own residual updates, without relying on a predefined safety direction. We compare our method with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models, including Qwen3 models from 0.6B to 32B. Results show our method improves jailbreak robustness while maintaining low over-refusal on benign safety-sensitive requests with modest inference overhead.
Large language models (LLMs) have shown strong capabilities across diverse applications, making both safety and utility essential for reliable deployment. Jailbreak attacks use diverse forms of adversarial prompting to bypass alignment safeguards and induce LLMs to generate harmful responses. Some representation-based defenses regulate internal activations using predefined steering directions or learned safety regions. However, adapting intervention to each input without a predefined representational target while limiting over-refusal remains a challenge. In this paper, we propose Target-free Entropy-guided Residual Modulation (TERM), an inference-time defense that uses prompt-calibrated attention entropy deviations to guide adaptive modulation of the model's own residual updates. Attention patterns vary across prompts and decoding steps, making raw entropy values difficult to interpret as a consistent signal for intervention. We therefore calibrate attention entropy against each prompt's own prefill baseline to establish an input-specific reference for reducing unnecessary intervention on benign requests. Adaptive scale normalization and temporal accumulation then account for differences in deviation scale and persistence to determine intervention strength. The resulting signal controls bounded scaling of the model's own residual updates, without relying on a predefined safety direction. We compare our method with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models, including Qwen3 models from 0.6B to 32B. Results show our method improves jailbreak robustness while maintaining low over-refusal on benign safety-sensitive requests with modest inference overhead.
Autonomous language-model agents can drive industrial controllers over fieldbus protocols, probing a plant at machine speed and in coordinated swarms. We present Talos, an inline gate that judges every agent request to a controller in milliseconds and routes agents it flags as hostile to a device-faithful decoy -- a mirror twin of the real plant -- where their attack continues harmlessly and is read for intelligence. We evaluate the full pipeline (detect -> quarantine -> mirror -> extract) against 2,400 autonomous attacker agents, 600 in each of four tiers that escalate the attacker's effort to stay hidden, from no concealment to an explicit non-disclosure rule and a trained cover story. No write reached the real plant in any tier: the gate denies the damage path and a device-level envelope check backstops it, while unmapped registers are denied without exception and the gate still permits safe-direction stop commands. Decisions are model-driven in every tier with no inference timeouts, closing a common objection to small on-device judges. Beyond blocking, the pipeline reads the attacker. We find a behavioral structure along two axes -- how persistently an agent attacks and how much it discloses -- and show that the system has a capture channel for three of the four resulting regions: hardened, silent attackers escalate most and are read on the mirror; compliant, talkative attackers volunteer their identity to a diagnostic form before they are ever quarantined; only the paranoid, low-yield region resists both. Attacker identity fields are self-authored cover stories, but involuntary tells -- the self-reported toolchain fingerprint and the attacker's own list of targeted registers -- accumulate across the swarm into a high-confidence profile drawn from agents under explicit orders to stay silent. The gate, mirror, and harness are released under Apache 2.0 with a synthetic device family.
An AI assistant writes small programs, and a gatekeeper must decide, from the code alone, whether each one may run. Suppose the rules are fixed: never perform a forbidden action, never let the output depend on a secret, always answer eventually. Can two rulebooks, one listing what is allowed and one listing what is not, both be complete, if every rule must be backed by evidence that a machine can check in finite time? We prove that they cannot. Checkable evidence can only establish what a finite record settles, and ``never'' and ``eventually'' are never settled by a finite record. For prohibitions the failure is one-sided: every violation can be caught by watching, but no safe program can be cleared for good. For confidentiality and responsiveness, neither side can be. We characterise exactly which rules about visible behaviour evidence can confirm, and we show that any sound rulebook of $n$ bits omits a compliant program of about $n$ bits. A time limit makes everything checkable, but only by running the program for essentially the whole limit, so a finite machine is not a feasible one. Finally, we show what a gatekeeper can still guarantee: not a verdict on every program, but a checkable list that contains everything it must allow and nothing it must refuse. Simulations on 2.2 million random programs and on self-referential adversaries show that the obstruction is adversarial rather than typical, and that the one error a longer test never removes is the liveness part of confidentiality.
The effectiveness of a self-improving system relies heavily on the quality of its verifier, which judges the outcome of each improvement round. However, the verifiers of most tasks are written by hand, which is expensive and does not scale to tasks that come without one. We propose Veritas, a harness that generates verifiers automatically. Unlike general-purpose agent harnesses, Veritas adds three mechanisms for verifier generation: an evidence header stating what the verifier will observe, a library of domain rules that tasks leave unstated, and authoring principles, partly refined by a self-improvement loop. We evaluate on six benchmarks: four scored by execution, whose official verifiers we first certify, and two that compare pairs of responses. With the model fixed, the verifiers Veritas generates agree better with the official verdicts than those of seven other harnesses on all four execution-scored benchmarks. On the two pairwise benchmarks they are at or near the best. They also agree better than those of the vendor's own harness for two of three other model families. Its mechanisms also improve 9 of 12 harness-benchmark pairs scored by execution across three harnesses.
Modern large language models (LLMs) are trained on massive, largely unfiltered datasets, including content scraped from nearly every accessible website and user inputs. As a result, LLMs often memorize and reproduce personally sensitive information (PSI) such as birth dates, phone numbers, and home addresses. This leads to significant privacy risks, particularly for high-profile individuals such as executives, politicians, and judges. Existing mitigations largely rely on machine unlearning. However, these methods often remove more information than needed, degrade model utility and safety, and are highly vulnerable to attacks. This paper presents Whiteout, a practical tool that, upon requests by individuals, prevents LLMs from regurgitating their genuine PSIs, by overwriting them using precise and carefully designed obfuscation samples. We evaluate Whiteout on modern LLMs of varying sizes and makers, including a widely-used OpenAI model. Results show that Whiteout effectively prevents disclosure of the targeted PSIs, has negligible impact on model utility and safety, and outperforms existing alternatives. We also test Whiteout against a wide range of countermeasures, from black-box attacks like jailbreaking to white-box adaptive attacks like relearning and quantization. Finally, we conclude with a discussion on the security and ethical implications of Whiteout.
Watermarking schemes for large language model output are evaluated almost exclusively on English text using each scheme's detection threshold and a narrow set of quality measurements. Multilingual deployment exposes evaluation-design choices that are inconsequential on English but determine conclusions cross-lingually. We propose an evaluation framework with four components: detection thresholds calibrated empirically per deployment context, a threshold-independent companion measurement that distinguishes calibration failures from detection failures, three disjoint quality measurement paradigms (distributional, paired-semantic, and reference-perplexity), and a generalized-entropy decomposition of cross-language disparity over a typological family partition. Applied to six watermarking schemes, three open-weight generators, eleven languages spanning four scripts and eight typological families, and both base and instruction-tuned regimes, the framework reveals failure modes that single-language, single-paradigm evaluation cannot surface. Against a shape-matched null, detection disparity is predominantly between-family on the typological partition, indicating that detection gaps are structural to language properties rather than idiosyncratic to particular languages; quality disparity is large but not reliably family-structured.
As LLM-based agents increasingly browse the web on users' behalf, a natural question arises: can websites passively identify which underlying model powers an agent? Doing so would represent a significant security risk, enabling targeted attacks tailored to known model vulnerabilities. Across 14 frontier LLMs and four web environments spanning information retrieval and shopping tasks, we show that an agent's actions and interaction timings, captured via a passive JavaScript tracker, are sufficient to identify the underlying model with up to 96\% F1 within a fixed harness. Although fingerprints are harness-specific, with cross-harness transfer falling to near chance, pooling traces across harnesses recovers model identification at 0.74 to 0.81 Macro-F1. We further show that strong classifiers can be trained from roughly one third of the available traces, and that identification improves steadily as more events are observed. Injecting randomised timing delays between actions substantially degrades classifier performance, but does not provide robust protection: a classifier retrained on delayed traces largely recovers performance. We release our code and a labelled corpus of agent traces at https://github.com/KabakaWilliam/known_actions.
Long-context LLMs can infer objectives that are not explicitly stated. This capability is useful for reasoning over documents, code, and tool traces, but it also introduces a safety risk. Harmful intent can be distributed across a context and become apparent only after the model combines the relevant information. Existing safety evaluations largely focus on explicitly stated harmful requests and therefore do not fully capture this failure mode. We introduce compositional reasoning attacks, a long-context attack setting in which harmful requests are decomposed into semantically incomplete fragments and embedded in long contexts. The final query is neutral, and the model must retrieve and combine the fragments to infer the implied harmful objective. Empirically, we show that models generally refuse harmful requests when they are explicitly stated. However, refusal rates drop when the same objectives must be reconstructed from fragments, and these failures often become more pronounced in longer contexts. Our analyses suggest that this gap cannot be explained by retrieval difficulties alone. Models can often reconstruct the harmful objective but do not reliably refuse it. Increasing inference-time reasoning effort improves refusal, but at a higher computational cost, with unsafe responses still occurring. Our results reveal a long-context safety gap: Models are better at refusing harmful requests they see than harmful objectives they infer.