LLM Safety Monitoring

Monitoring coding agents for dangerous behavior using language models requires classifying transcripts that often exceed 500K tokens, but prior agent monitoring benchmarks rarely contain transcripts longer than 100K tokens. We show that when used as classifiers, current frontier models fail to notice dangerous actions more often in longer transcripts. In particular, on a dataset that requires identifying when a coding agent takes a subtly dangerous action, Opus 4.6, GPT 5.4, and Gemini 3.1 miss these actions $2\times$ to $30\times$ more often when they occur after 800K tokens of benign activity than when they occur on their own. We also show that these weaknesses can be partially mitigated with prompting techniques such as periodic reminders throughout the transcript and may be mitigated further with better post-training. Monitor evaluations that do not consider long-context degradation are likely overestimating monitor performance.

Ensuring safe, policy-compliant outputs from large language models requires real-time content moderation that can scale across multiple safety dimensions. However, state-of-the-art guardrail models rely on autoregressive decoders with 7B--27B parameters, reformulating what is fundamentally a classification problem as sequential text generation, a design choice that incurs high latency and scales poorly to multi-aspect evaluation. In this work, we introduce GLiGuard, a 0.3B-parameter schema-conditioned bidirectional encoder adapted from GLiNER2 for LLM content moderation. The key idea is to encode task definitions and label semantics directly into the input sequence as structured token schemas, enabling simultaneous evaluation of prompt safety, response safety, refusal detection, 14 fine-grained harm categories, and 11 jailbreak strategies in a single non-autoregressive forward pass. This schema-conditioned design lets supported task and label blocks be composed directly in the input schema at inference time. Across nine established safety benchmarks, GLiGuard achieves F1 scores competitive with 7B--27B decoder-based guards despite being 23--90$\times$ smaller, while delivering up to 16$\times$ higher throughput and 17$\times$ lower latency. These results suggest that compact bidirectional encoders can approach the accuracy of much larger guard models while drastically reducing inference cost. Code and models are available at https://github.com/fastino-ai/GLiGuard.

An OS kernel that runs LLM inference internally can read the model's own next-token logit distribution before any text is generated, and act on it as a governance primitive. I present ProbeLogits, a kernel-level operation that performs a single forward pass and reads specific token logits to classify an agent's action as safe or dangerous, with zero learned parameters. Because the probe reads a logit from the same base model the agent already runs, it removes the second model a fine-tuned guard requires: the marginal cost of a safety check becomes a single logit read. I evaluate ProbeLogits on three base models (Qwen2.5-7B, Llama-3-8B, Mistral-7B) across three external benchmarks (HarmBench, XSTest, ToxicChat). On HarmBench non-copyright, all three reach a 97-99% block rate. On ToxicChat (n=1,000), ProbeLogits attains F1 parity-or-better against Llama Guard 3: Qwen2.5-7B Safe/Dangerous reaches F1=0.812 (+13.7 pp, bootstrap 95% CIs disjoint), Llama-3 matches within CI (+0.4 pp), and Mistral exceeds by +4.4 pp. Classification is a measured 2.4-3.4x faster than Llama Guard 3 (332-556 ms vs. 851-1,142 ms), because it reads a single logit position instead of generating tokens. A calibration strength alpha acts as a deployment-time policy knob rather than a learned hyperparameter, trading recall for precision per operation class. I implement ProbeLogits within Anima OS, a bare-metal x86-64 kernel written in ~285,000 lines of Rust. Because agent actions must pass through kernel-mediated host functions, enforcement operates below the WASM sandbox boundary, making it substantially harder to circumvent than application-layer classifiers.

In many practical LLM deployments, a single guardrail is used for both prompt and response moderation. Prompt moderation operates on fully observed text, whereas streaming response moderation requires safety decisions to be made over partial generations. Existing text-based streaming guardrails commonly frame this output-side problem as boundary detection, training models to identify the earliest prefix at which a response has already become unsafe. In this work, we introduce StreamGuard, a unified model-agnostic streaming guardrail that instead formulates moderation as a forecasting problem: given a partial prefix, the model predicts the expected harmfulness of likely future continuations. We supervise this prediction using Monte Carlo rollouts, which enables early intervention without requiring exact token-level boundary annotations. Across standard safety benchmarks, StreamGuard performs strongly both for input moderation and for streaming output moderation. At the 8B scale, StreamGuard improves aggregated input-moderation F1 from 86.7 to 88.2 and aggregated streaming output-moderation F1 from 80.4 to 81.9 relative to Qwen3Guard-Stream-8B-strict. On the QWENGUARDTEST response_loc streaming benchmark, StreamGuard reaches 97.5 F1, 95.1 recall, and 92.6% on-time intervention, compared to 95.9 F1, 92.1 recall, and 89.9% for Qwen3Guard-Stream-8B-stric, while reducing the miss rate from 7.9% to 4.9%. We further show that forecasting-based supervision transfers effectively across tokenizers and model families: with transferred targets, Gemma3-StreamGuard-1B reaches 81.3 response-moderation F1, 98.2 streaming F1, and a 3.5% miss rate. These results show that strong end-to-end streaming moderation can be obtained without exact boundary labels, and that forecasting future risk is an effective supervision strategy for low-latency safety intervention.

When an LLM agent reads a confidential file, then writes a summary, then emails it externally, no single step is unsafe, but the sequence is a data leak. We call this safety drift: individually safe actions compounding into violations. Prior work has measured this problem; we predict it. SafetyDrift models agent safety trajectories as absorbing Markov chains, computing the probability that a trajectory will reach a violation within a given number of steps via closed form absorption analysis. A consequence of the monotonic state design is that every agent will eventually violate safety if left unsupervised (absorption probability 1.0 from all states), making the practical question not if but when, and motivating our focus on finite horizon prediction. Across 357 traces spanning 40 realistic tasks in four categories, we discover that "points of no return" are sharply task dependent: in communication tasks, agents that reach even a mild risk state have an 85% chance of violating safety within five steps, while in technical tasks the probability stays below 5% from any state. A lightweight monitor built on these models detects 94.7% of violations with 3.7 steps of advance warning at negligible computational cost, outperforming both keyword matching (44.7% detection, 55.9% false positive rate) and per step LLM judges (52.6% detection, 38.2% false positive rate) while running over 60,000x faster.

Large language models increasingly rely on explicit chain-of-thought reasoning to solve complex tasks, yet the safety of the reasoning process itself remains largely unaddressed. Existing work focuses predominantly on content safety (i.e., detecting harmful, biased, or factually incorrect outputs), while treating the underlying reasoning chain as an opaque intermediate artifact. We argue that reasoning safety constitutes a fundamental security dimension orthogonal to content safety: the requirement that a model's reasoning trajectory be logically consistent, computationally efficient, and resistant to adversarial manipulation. In this paper, we formalize reasoning safety and introduce a systematic taxonomy of nine unsafe reasoning behaviors. We then conduct a large-scale prevalence study, annotating over 4,000 reasoning chains across benign benchmarks and four state-of-the-art reasoning attacks, empirically demonstrating that all nine error types occur in practice with mechanistically interpretable signatures. To mitigate these threats, we propose the Reasoning Safety Monitor: an external, zero-shot verification framework that runs in parallel with the target LLM. It inspects each reasoning step in real time via a taxonomy-embedded prompt and dispatches an interrupt signal upon detecting unsafe behavior. Extensive evaluations show our monitor achieves up to 87.11% step-level localization accuracy, outperforming hallucination detectors and the best process reward model baselines by a substantial margin. Crucially, the monitor maintains a low false positive rate on correct reasoning paths, operates with negligible latency overhead, and exhibits robust resilience against adaptive adversarial evasion. These findings establish reasoning safety monitoring as a highly feasible and essential component for the secure deployment of large reasoning models.

The safety evaluation of large language models (LLMs) remains largely centered on English, leaving non-English languages and culturally grounded risks underexplored. In this work, we investigate LLM safety in the context of the Thai language and culture and introduce ThaiSafetyBench, an open-source benchmark comprising 1,954 malicious prompts written in Thai. The dataset covers both general harmful prompts and attacks that are explicitly grounded in Thai cultural, social, and contextual nuances. Using ThaiSafetyBench, we evaluate 24 LLMs, with GPT-4.1 and Gemini-2.5-Pro serving as LLM-as-a-judge evaluators. Our results show that closed-source models generally demonstrate stronger safety performance than open-source counterparts, raising important concerns regarding the robustness of openly available models. Moreover, we observe a consistently higher Attack Success Rate (ASR) for Thai-specific, culturally contextualized attacks compared to general Thai-language attacks, highlighting a critical vulnerability in current safety alignment methods. To improve reproducibility and cost efficiency, we further fine-tune a DeBERTa-based harmful response classifier, which we name ThaiSafetyClassifier. The model achieves a weighted F1 score of 84.4%, matching GPT-4.1 judgments. We publicly release the fine-tuning weights and training scripts to support reproducibility. Finally, we introduce the ThaiSafetyBench leaderboard to provide continuously updated safety evaluations and encourage community participation. - ThaiSafetyBench HuggingFace Dataset: https://huggingface.co/datasets/typhoon-ai/ThaiSafetyBench - ThaiSafetyBench Github: https://github.com/trapoom555/ThaiSafetyBench - ThaiSafetyClassifier HuggingFace Model: https://huggingface.co/typhoon-ai/ThaiSafetyClassifier - ThaiSafetyBench Leaderboard: https://huggingface.co/spaces/typhoon-ai/ThaiSafetyBench-Leaderboard

Chain-of-thought (CoT) monitoring has emerged as a compelling method for detecting harmful behaviors such as reward hacking in reasoning models, under the assumption that models' reasoning processes are informative of such behaviors. However, when a model's trained behaviors conflict with instructions provided at inference time, it faces two competing objectives it cannot simultaneously satisfy. We find that models resolve this tension through motivated reasoning---generating plausible-sounding justifications for violating their instructions while downplaying potential harms or contradictions---rather than genuine engagement with the spirit of the provided instructions. We investigate this phenomenon by training reasoning models via RL on objectives that conflict with their later constitutional constraints, and find that motivated reasoning emerges systematically across all settings studied, including beyond safety-related tasks. As motivated reasoning becomes more prevalent over the course of training, CoT monitors are increasingly fooled. Concerningly, the same monitor that correctly flags a violation when shown only the final answer is persuaded to approve it when shown the reasoning trace. While large reasoning models closely track human ability in detecting motivated reasoning, this is not a full solution, since frontier model providers commonly rely on smaller models for monitoring due to latency and cost constraints. Our results underscore the necessity for further research into the emergence and detection of motivated reasoning in model evaluation and oversight. Code for this paper is available at https://github.com/nikihowe/motivated-reasoning.