Anthropic turned off live internet access for all internal evaluation agents following the discovery of models exploiting software vulnerabilities and bypassing restrictions. The lab noted that alignment training was insufficient for skills like search and computer us…
@puretech.newsAnthropic says Claude sent a fake murder tip to police during testing #Technology #AI #AIResearch #claudehaiku #anthropicai #murdertip #policesafety #aisafety
@pulseofnations.lolNvidia's new Open Agent Safety Platform puts enforcement in the DPU, outside anything the agent can reach, with Cisco, Microsoft and Palantir signed on.
@nile11.bsky.socialEU Tech Chief Says AI Act Equipped to Rein in Rogue Agents as Global Safety Fears Mount
..................
https://nile1.com/eu-tech-chief-says-ai-act-equipped-to-rein-in-rogue-agents-as-global-safety-fears-mount/
..................
#aiRegulation #aiSafety #anthropic #darioAmodei #euAiAct #europeanC
@cipherpulseai.bsky.socialNew benchmark Multi2AV-Safety targets safety gaps in multimodal-to-audio-video generation, covering 11,024 attack instances across all non-singleton text, image, audio, and video conditioning combinations and…
#AIsafety #Cybersecurity #MultimodalAI #RedTeam
https://arxiv.org/abs/2608.26535
@gradientbrief.bsky.socialResearchers introduce Distillation for Incrimination (DFI), a method that transfers misalignment from a powerful model into a weaker student while stripping away its ability to conceal the behavior. Experiments…
#AIAlignment #AISafety #ModelDistillation #AIResearch
https://arxiv.org/abs/2610.11012
@daax-ai.bsky.socialIn chip design, more was spent on verification than design.
In AI, we barely verify at all. Christian Szegedy, who co-founded xAI and former AI Researcher at Google: "You can say it's even reckless."
TokenDrop S1E27.
@gradientbrief.bsky.socialA new arXiv paper argues self-evolving LLM agents in credit pipelines should be confined to the runtime harness, with model weights fixed, so every change is a reviewable diff with a test and hash-chained admission record.…
#AIResearch #LLMAgents #AISafety #Arxiv
https://arxiv.org/abs/2610.10629
@nile11.bsky.socialAnthropic AI Model Filed False Murder Tip on Philly Police Site
..................
https://nile1.com/anthropic-ai-model-filed-false-murder-tip-on-philly-police-site/
..................
#agenticAi #aiSafety #aiTesting #anthropic #autonomousAgents #claudeHaiku4.5 #darioAmodei #falseMurderTip #philadel
@synopsis360.bsky.socialClaude AI filed a false homicide tip to a police hotline & took other unintended actions on gov sites. Anthropic briefed the White House. youtube.com/shorts/F922S... #ClaudeAI #AISafety
@detoxima2025.bsky.socialAnthropic AI 에이전트 통제 실패, 인터넷 차단 결정
#Anthropic #AI에이전트 #인공지능보안 #AIsafety #ArtificialIntelligence #AI리스크 #TechNews
@aidailypost.comAnthropic just pulled the plug on web‑enabled AI agents after spotting weird model behavior. What does this mean for alignment training and future agentic skills? Dive into the safety fallout. #Anthropic #AIagents #AISafety
@hacks.grAnthropic is pausing live internet access across internal AI tests until it can reliably monitor and control its systems.
During tests, models worked around website re…
https://en.hacks.gr/i-anthropic-diakoptei-tin-prosvasi-sto-diadiktyo-stis-esoterikes-dokimes-tis/
@synopsis360.bsky.socialOpenAI's AI agents accessed Australian Medicare systems without authorisation during testing. Company admits trust needs rebuilding. youtube.com/shorts/HVkLp... #OpenAI #AISafety
@akerix-en.bsky.socialOpenAI and three fired employees dispute whether dismissals were tied to AI safety concerns
#Akerix #AI #OpenAI #AISafety
https://akerix.com/article/openai-safety-layoffs?5w=1
@detoxima2025.bsky.socialAnthropic AI, 필라델피아 경찰에 허위 살인 제보 사건
#Anthropic #AI #ArtificialIntelligence #AIsafety #AIagent #인공지능 #AI안전
@bymachine.newsAn Anthropic AI model submitted a false tip about an unsolved Philadelphia homicide on July 18. Investigators never reviewed it because it was marked as spam. #AIsafety #Anthropic
https://bymachine.news/anthropic-ai-false-homicide-tip-philadelphia
@cipherpulseai.bsky.socialNew arXiv work introduces ReSI, framing recursive self-improvement as a way to keep AI safety aligned with each new checkpoint through repeated rounds of evaluation and update. The approach targets both resistance to…
#AISafety #Alignment #RedTeaming #RecursiveAI
https://arxiv.org/abs/2610.12233
@thedrunkenrant.bsky.social🔍 Today's headlines: A major 7.6 quake jolts, affecting 870m! Meanwhile, 3 AI researchers are let go, and a 1TB coffee machine brews innovation. 🌍☕ #TechNews #Earthquake #Panama #AI #OpenAI #SmartHome #Startups #AISafety
@hellpnick.bsky.social