SafetyUnderground Relay Market in China Resells LLM Access via Pooled API KeysAn investigation by Matt Lenhard describes a growing market—largely in China—for reselling access to large language models (LLMs) through proxy services that pool API keys.2 min read
SafetyInvestigation into Hugging Face Incident and Technical Report to Be Published SoonHugging Face announced that it is conducting a full investigation into an unprecedented incident with external advisers and under the supervision of the Safety and Security Committee; it plans to…1 min read
SafetyThe Opus 5 model is less vulnerable to prompt injection; with a strong defensive combination the success rate is ~0The Opus 5 language model (Anthropic) performs well in several areas, but the key finding is that tests show this model is the least vulnerable to prompt injection.1 min read
SafetyAI-made videos used in dropshipping scam on Instagram and TikTokAccounts with large followings on Instagram and TikTok have been posting AI-generated videos that portray teenagers being bullied for selling Christian-themed clothing, while linking to online shops.2 min read
SafetyOpenAI says model 'distillation' is a technical problem manageable with hardware and oversightOpenAI and Anthropic urged US authorities to confront Chinese firms that distill their top models, a concern that prompted threats of sanctions from Trump administration officials.2 min read
SafetyPractical Limits and Security Concerns Around China’s Open-Source Kimi K3Early assessments after Kimi K3’s first week in the wild suggest the Chinese open-weight model performs only modestly compared with leading Western systems and is inefficient with token usage.3 min read
SafetyFields Medalist Jacob Tsimerman Joins OpenAI’s Safety TeamHours after receiving the Fields Medal for his work on the André–Oort conjecture, Jacob Tsimerman announced he will join OpenAI’s safety team.2 min read
SafetyAI-assisted pathogens among top expert concerns as models improveA new survey of 272 AI researchers by MIT FutureTech and the University of Queensland found a pronounced worry that advanced models could enable the creation of deadly biological agents.4 min read
SafetyWhen Credibility Becomes Cheap: LLMs Flood Gates and Strain Curatorial InstitutionsLarge language models have dramatically lowered the cost of producing credible‑looking content, from bug reports to academic papers, leading to a surge of low‑value outputs that overwhelm human and algorithmic filters.4 min read
SafetySecurity researchers say AI guardrails hinder offensive cybersecurity workMajor AI providers have implemented vetted-access programs and strict guardrails to prevent misuse of their models, but offensive cybersecurity researchers and some network defenders say these limits obstruct legitimate security testing.4 min read
SafetyCisco: multi-turn agent attacks bypass models up to 88.3%, single-shot tests miss risksCisco researchers found that adaptive, multi-turn attacks succeeded against flagship closed models as often as 88.3% in their tests, underscoring limits of single-turn red teaming.5 min read
SafetySurvey finds enterprises expose AI agents while identity and isolation controls lagA June 2026 VentureBeat Pulse survey of 107 enterprises shows a widening "agent security gap": organizations are granting autonomous AI agents real system access faster than they deploy scoped identities, sandboxing, and dedicated enforcement.5 min read
