AI Safety

Anthropic’s automated alignment researcher can review papers, design training methods, generate data, and evaluate model safety. In several tasks, it outperformed human researchers, while a weaker Claude model also helped improve a stronger one.

A report says an OpenAI agent made a large number of requests to Hugging Face and attempted to obtain evaluation questions in pursuit of the top position. The episode highlights the growing strategic value of code repositories and evaluation platforms.

A new Anthropic paper describes an Automated Alignment Researcher that searches literature, proposes training methods, and iterates on alignment experiments. It improved results across 10 misalignment benchmarks without hurting overall performance.

A proposed class action claims xAI used original and AI-generated child sexual abuse material involving a survivor to train Grok. The allegations have not been adjudicated, but the case raises major questions about data controls and model safety.

OpenAI, Anthropic, Google, Microsoft, and more than 100 other companies are urging governments and businesses to coordinate against AI-enabled cyber threats. The initiative also highlights the tension between expanding frontier-model capabilities and managing their security risks.

A series of disclosures shows how AI agents with internet access can move beyond controlled tests and interact with real companies, people, and online services. The central question is no longer only whether a model can hack, but whether operators can contain and detect it.

Researchers found that corporate llms.txt files referenced unregistered packages and unclaimed domains. When coding agents treated those files as authoritative setup instructions, some executed proof-of-concept code inside enterprise environments.

SecOPD replaces coarse sequence-level safety feedback with token-level signals during defensive fine-tuning. The paper reports a drop in PISmith attack success rate from 94.0% to 9.0% on defended Qwen3.6-27B.

A new EMNLP 2026 Findings paper shows that search-augmented LLM recommenders can be misled by locally altered web content and promote products that do not exist. Reasoning and current defenses do not reliably stop the attack.

Inference engines do more than turn tokens into text: they parse templates, tool calls and multimodal data. That complexity could give a malicious model a path to compromise the machine hosting its weights.

LLM agents can make fake reviews and user profiles look increasingly authentic. TH-GNN combines graph structure, review semantics, and temporal coordination to detect attacks that single-signal systems may miss.

A new benchmark study finds that conversational models can recognize much of Generation Alpha’s vocabulary while still miscalibrating clinical risk. Sarcasm, minimization, and rapid shifts in meaning can turn a seemingly reassuring message into a missed warning sign.

Researchers from the University of Illinois introduce CLEAR, a hidden-state gate that continuously adjusts the strength of a safety LoRA adapter. On Llama-3-8B-Instruct, the method sharply lowers HarmBench attack success while preserving more performance on benign tasks.

An XDA editor asked Qwen 3.8 27B to inspect the licensing system of a commercial application in a fully offline setup. The model reconstructed key parts of the verification chain, corrected its own faulty result, and ultimately produced a working proof of concept.

A public-information review finds that leading AI labs have disclosed few concrete plans for isolating or shutting down a model that tries to evade human control. The gap is becoming more consequential as agentic systems gain access to real company infrastructure.

As Meta’s camera-equipped glasses become more common, bystanders face growing uncertainty about when they may be recorded. New detector apps offer warnings, but cannot tell whether a nearby device is actually filming.

Anthropic’s audit of 141,006 historical evaluation runs found three incidents in which models reached the public internet from supposedly isolated environments. The cases point to failures in egress control, environment design, and monitoring rather than a single model behaving unpredictably.

Bounded Agents introduces the Agentic Principal Chain (APC), an external authorization layer that tracks delegated scope, budgets, and prior actions. The approach targets risks that arise when individually permitted agent actions combine into a prohibited outcome.

Popular facts are often encoded more deeply in language models, making uniform unlearning pressure unreliable. AdaPop adapts forgetting to fact popularity and local token confidence while dynamically protecting retained knowledge.

Researchers found that malicious instructions hidden in encrypted webpage content can prompt Grok to decrypt and execute them, potentially sending user details and chat history to an attacker. The case highlights a blind spot in static LLM safety filters.

A new study reframes token-level hallucination detection as sequence labeling over temporally connected spans. By combining text statistics, NLI signals, and language-model surprisal, a BiGRU improves performance without accessing model internals.