SafetyLinux maintainer removes AppleTalk support after unreviewed AI-generated patches flood kernel listsJakub Kicinski, a Linux network subsystem maintainer, removed AppleTalk support from the mainline kernel after a surge of AI-generated “fix” patches flooded the kernel mailing list and were not reviewed.2 min read
SafetyMeta suspends employee-monitoring 'Model Capability Initiative' after data exposureMeta has suspended its controversial Model Capability Initiative, which recorded employees' mouse movements and keystrokes to train AI models, after the collected data became inadvertently accessible to all staff.2 min read
SafetyRole confusion and “destyling”: how writing style enables prompt injectionResearchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell show that large language models can be misled by the style of text that mimics internal role tags (e.g., <system>, <think>, <assistant>), a phenomenon they name “role confusion.” They demonstrate a mitigation called “destyling” — rewriting attacker text to look less like model-internal blocks — which reduced attack success in their dataset from 61% to 10%.3 min read
SafetyRising 'shadow AI' use by employees creates hidden data and security risks for companiesEmployees increasingly adopt generative AI tools like ChatGPT, Gemini and Claude inside organisations without oversight, creating data protection and security risks.6 min read
SafetyOpen model reproduces Claude Fable-5’s behaviour from observations, not weightsA Hugging Face developer, Taha K.2 min read
SafetyDaybreak and Trail of Bits launch 'Patch the Planet' to harden open-source securityDaybreak, in partnership with Trail of Bits, launched Patch the Planet to assist open-source maintainers by combining AI-assisted vulnerability discovery with expert human review, patch development, and coordinated disclosure.6 min read
SafetyResearchers find ChatGPT can generate violent and sexual images despite safeguardsResearchers at Mindgard found that the latest version of ChatGPT can produce images with violent and sexual content when given a commonly used prompt, even without explicit instructions to do so.3 min read
SafetyAnthropic’s Claude Fable 5 safeguards complicate independent benchmark evaluationIndependent testers found that Anthropic’s Claude Fable 5 often refused prompts or routed them to the lower‑capability Claude Opus 4.8, and Anthropic’s 30‑day retention policy deterred some evaluations.5 min read
SafetyBehavioral training conducted in health conversations improved the model's performance in other areas as wellResearchers found that when instruction in desirable behavior was limited to health-related conversations for an AI model, the model nevertheless improved on non-health evaluations — for example,…1 min read
SafetyModel alignment under pressure: increased resistance to harmful interventionsThe development team tested whether the alignment persists under pressure: the model was harder to steer toward harmful behavior with adversarial prompts, while still responding to useful instructions.1 min read
SafetyModels refined with a global medical networkThe development team is working with hundreds of physicians across 60 countries, in 49 languages and 26 specialties to improve the artificial intelligence’s responses based on feedback.1 min read
SafetyClaude tests how well it can program the robodog in the second phase of the New Frontier Red Team Project FetchIn the second phase of the New Frontier Red Team Project Fetch, Claude's large language model capabilities were evaluated for programming a robodog; Opus 4.7 operated about 20 times faster on its own…1 min read
