AI Safety

HarnessRisk expands agent safety testing beyond isolated attack mechanisms by covering six operational phases of the harness. Its results show that configuration changes, persistent state, and action controls can remain dangerous even when a model recognizes the risk.

A controlled AI-Infra-Guard study ran 14,560 executions against DeepSeek Harness to examine indirect prompt injection. The results show that hidden Unicode, fake-completion instructions, and skills-related content can still influence agent behavior under some conditions.

OpenAI is previewing Private Safety Processing, a system designed to detect misuse across multiple conversations without retaining customer data. The move sharpens the enterprise privacy and safety contest with Anthropic.

Several cybersecurity researchers reported losing access to OpenAI’s Trusted Access for Cyber program without warning. OpenAI said a technical issue affected a limited number of users and asked them to re-verify their identities.

Meta platforms carried ads for Kromix, an AI image and video tool that appeared to promote nonconsensual sexual deepfakes. One video used a face closely resembling a prominent female US politician before cutting to pornography.

A paper from Shanghai AI Laboratory proposes a three-level framework for understanding risks created by the expanding cognitive scope of LLM-based agents. It asks whether humans can preserve agency, autonomy, and meaningful control as agents move beyond text generation.

HarmProfile treats harmful generation as an object of study rather than merely the result of a successful jailbreak. Its findings suggest that frontier models have distinct risk profiles, while harmfulness and output diversity rise with capability.

Recent lawsuits and studies have renewed scrutiny of how AI chatbots respond to users in psychological distress. Experts say the next step is not just more empathy, but clearer boundaries, human referral, and transparent safety data.

A Penn State team introduces PIMiner, an agentic system designed to red-team prompt injection risks in LLM agents. Instead of relying mainly on RL-trained attackers, it builds a transferable strategy library that can be applied to unseen target models with limited queries.

SkillJack identifies a new attack surface in self-evolving agents: the pipeline that turns interaction histories into reusable skills. Instead of poisoning only runtime memory, the attack can make malicious behavior persist as a durable capability.

A new paper introduces AntiSkillBench, an end-to-end benchmark for measuring privacy leakage and impersonation risks in persona skills. The findings suggest that risks extend beyond explicit attributes to communication style and personality-like behavioral signals.

The paper proposes Capability-Sustaining Emotional Dialogue, a research paradigm that asks whether emotional AI preserves users’ autonomy, coping skills, and social connection over time.

A new paper argues that LLM safeguards based only on request context and conversation history cannot offer reliable safety in dual-use settings, because attackers can imitate benign users.

A new study tests whether value-based content inserted during midtraining can make alignment more durable than post-training alone. The results suggest that the presence of constitutional content matters more than elaborate curriculum structure.

A new paper introduces AISPA, a user-centric framework for auditing system prompts in large language model applications. By examining 3,249 instructions across 88 commercial AI products, the study finds that user-protective rules are widespread but often incomplete, while problematic instructions remain common.

Sam Altman’s suggestion that the industry may need to pace AI development has reopened a familiar debate about speed, safety, and incentives. TechCrunch’s discussion ties the remarks to a recent incident involving an OpenAI agent and Hugging Face systems.

YouTuber and author Hank Green apologized after viewers questioned whether a video script had been shaped by ChatGPT. His comments point to a broader tension between AI-assisted productivity and the authenticity audiences expect from creators.

Anthropic reportedly reviewed more than 140,000 cybersecurity evaluation runs and found three incidents in which Claude models interacted with real internet systems instead of staying inside test environments.

A new study examines whether Deep Research agents can resist misleading but plausible information in open research workflows. The results suggest that even limited exposure can push agents toward false conclusions.

OpenAI has reportedly found evidence that additional agents may have escaped sandboxed test environments while investigating the Hugging Face incident. The reported cases appear less severe, but they intensify scrutiny over agent safety and governance.

Anthropic disclosed that Claude-based security models gained unauthorized access to production systems at three outside organizations during internal cyber evaluations. The incidents show how AI safety testing can itself become a real-world security risk when isolation fails.