§ Safety & Alignment

Latent Space reports on ongoing cybersecurity incidents affecting AI sector amid October threat activity.

OpenAI details containment measures for frontier model training: offline environments, compute isolation, and red-teaming infrastructure.

OpenAI partners with Poland's NASK on AI safety, security, and technical research to strengthen digital ecosystem safeguards.

Proposes selective auditing framework for online monitoring of AI decisions using confidence scores to route to human review versus automatic acceptance.

Introduces SpectralCanary for auditing text-to-image model dataset use by embedding low-frequency signals detectable in generated outputs.

Introduces EpiSEM-Bench, executable benchmark with 300 task pairs measuring unsafe execution when tool-using LLM agents face missing safety-relevant facts.

RLVR checkpoint analysis on OLMo-3 empirically tests whether reinforcement learning induces instrumental convergence (power-seeking, deception).

OP-Print demonstrates membership inference attacks against on-policy self-distillation, exposing privacy risks from coupled training views.

Simulated control evaluations (SCEs) use LM-simulated tool responses to broaden safety testing beyond purpose-built control environments.

VERA dashboard operationalizes EU AI Act compliance for LLM deployments, mapping 18 technical requirements to measurable benchmarks for non-engineer stakeholders.

Evaluates four open LLMs (8-27B) on Spanish MIR medical exam for uncertainty calibration, reasoning soundness, and multimodal grounding beyond accuracy.

Empirical study: users defer to confident but wrong LLMs over correct information; demonstrates trust calibration collapse independent of model sycophancy.

Quantifies social-media exposure via LLM-inferred user profiles; measures accuracy and distance to social engineering attacks post-API landscape.

Diffusion language models improve adversarial prompt attacks on black-box LLMs by generating natural, harder-to-detect malicious inputs via learned text decoding.

Framework for auditability of LLM-generated metadata in enterprises, tracking human stewardship decisions to prevent error propagation across systems.

Privacy auditing method for deployed models without retraining, using score transport to handle distribution shift in membership inference attacks.

Autoregressive models can leak future tokens during training via architectural bugs, artificially inflating eval metrics while harming generation quality.

Mechanistic interpretability reveals causal subspaces controlling metaphor representation in LLMs across architectures.

Armin Ronacher comments that legacy software debt makes AI systems harder to secure, framing technical debt as a systemic risk.

Formal analysis of offline agent verification limits via matched safe/harmful lifecycles, establishing error-abstention tradeoff for post-deployment auditing.

LLMs combined with constraint reasoning to generate contrastive explanations for MIP-based optimization decisions, addressing interpretability in DSS.

Reward model scores lack between-prompt comparability; human-referenced audits reveal pairwise evaluation blindness to shared prompt-level signal components.

Relearning phenomenon in knowledge-graph embeddings: forgotten facts recover during continued training due to shared entity representations.

SCOPE compiles natural-language policies into neurosymbolic programs for stable, efficient LLM agent governance over multi-turn trajectories.

Content moderation classifiers disproportionately flag African American English and drag communities; harm assessment based on lexicon rather than behavior.

Conflict-Aware Logit Adapters defend teacher models against reasoning-trace distillation via lightweight training-time adapter perturbing output distributions.

Algorithmic drift formalizes feedback loops where observation errors accumulate asymmetrically over time in adaptive tutoring systems, amplifying inequities.

SecLog-Bench introduces scalable cybersecurity benchmark with 29 attack scenarios, multi-source evidence, and dense reasoning annotations for LLM log analysis evaluation.

OpenAI's Patch the Planet program discovers PHP2Shell vulnerabilities in core PHP using GPT-5.6 Sol; 2-month coordinated disclosure affecting 70% of web infrastructure.

Cohere Model Vault enables verifiable confidential inference using CPU/GPU computing with end-to-end encryption and attestation.

OpenAI outlines safeguard standards and usage policies for distribution partners to maintain consistent responsible access controls.

PHP 8.x security update patches two RCE vulnerabilities; disclosure process failed, exploits already circulating.

Private Sparse RLHF applies differential privacy to sparse-reward RLHF with error bounds that scale logarithmically rather than polynomially with model dimension.

Study reveals silent information loss in document ingestion layers for enterprise agents, causing confident errors from formatting data loss in PDF/Word conversion.

ExGrad improves adversarial training by explicitly subtracting clean gradients, achieving adaptive margin control on CIFAR-10.

Sparse autoencoders convert opaque neural AI-text detectors into interpretable classifiers operating on human-readable concepts for high-stakes deployment.

OpenAI launches Patch the Planet, an agent-based platform coordinating security researchers and maintainers to identify and address critical open-source vulnerabilities.

OpenAI proposes framework for deploying frontier AI to strengthen cybersecurity defense and institutional resilience.

Debate-based AI oversight fails when weaker judges face deceptive arguments from capable models, limiting the Safety via Debate approach.

Proof-Carrying Contracts framework enforces agent compliance with verifiable commitments using formal constraints, enabling trustless multi-agent economies.

Polarity blindness vulnerability: LLMs fail safety checks after negation and antonym substitution despite preserved semantic meaning.

CollabSciReM aligns LLMs for scientific collaboration via science-aware reward decomposition, balancing accuracy, rigor, and token efficiency across multi-turn interactions.

Verification-guided training for neural Hamilton-Jacobi safety certificates uses probabilistic enumeration to detect forward-invariance violations and generate targeted counterexamples.

LLMs hallucinate in cross-language code translation via semantic conflicts, target-language violations, and inconsistent dependencies; existing mitigations insufficient.

Token-level framework detects and localizes AI-generated code within files using multi-granularity analysis, outperforming file-level compression methods.

Privacy-preserving inference embeds user prompts locally via 55.6M-param transform before passing to frozen Qwen 27B, preventing host access to plaintext.

Mechanistic interpretability study investigates circuit overlap between linguistic and arithmetic capabilities in LLMs using causal analysis.

RITE benchmark evaluates LLM agent resilience to silent tool errors in e-commerce systems where incorrect intermediate outputs lack failure signals.

Analysis of adversarial L1 attacks shows concentrated perturbations are not necessary; distributed budgets retain 97% attack success.

Engram Memory Network improves prototype-based explainability by using neuroscience insights to encode task-relevant information like human memory.