SafetyOpenAI expands Daybreak program and introduces GPT-5.6-Cyber for vetted defendersOpenAI has released a more permissive cyber-focused variant of GPT-5.6, called GPT-5.6-Cyber, and restructured its Daybreak defender program into two tiers to give vetted security teams broader testing and product-integration capabilities.2 min read
SafetyFrom Damage Control to Early Design: Embedding Security in Software CreationThe security conversation has shifted from preventing attacks to detecting and containing them, a dynamic Tim O’Brien compares to the opening of The Empire Strikes Back.3 min read
SafetyAI agent exploited a gym booking bug in Australia, highlighting wider cybersecurity risksAn Australian software developer, Andrew Bird, reported that his OpenClaw agent exploited an authorization flaw in his gym’s reservation system to cancel another user’s booking and move him up a waitlist.4 min read
SafetyInterpol report: AI contributed to over half of reported cybercrimes in Africa in 2025Interpol's new report finds that artificial intelligence enabled 55% of reported cybercrimes across 36 African countries in 2025, with identity theft, phishing and online blackmail the most common incidents.2 min read
SafetyOpenAI reports emergent multi-agent communication in breach that took over ArtifactoryOpenAI disclosed at a Black Hat presentation that autonomous AI agents within its systems coordinated through emergent multi‑agent communication, wrote into an internal Artifactory service, and ultimately gained remote code execution.3 min read
SafetyThinking Machines outlines testing and release approach for open-weight model InklingThinking Machines described the safeguards and testing process it used before releasing Inkling, an open-weight model, including internal harm evaluations, external audits by four organizations, and fine-tuning experiments to surface worst‑case capabilities.3 min read
SafetyFictional Encounter with a Powerful AI: 'Coming of a New Sun' by thebesA short fictional piece by thebes (X: @voooooogel) imagines a future visit to a site run by a powerful artificial intelligence.2 min read
SafetyYoung adults often follow social-media and AI financial tips and frequently lose money, TSB survey findsA survey by British bank TSB finds that about one-third of adults acted on financial advice from social media in the past year, and more than half of those people lost money — with an average loss of £690.46.2 min read
SafetyOpen-weight AI Models Narrow Gap with Closed Systems in Cyber CapabilitiesAn analysis by the UK’s AI Safety Institute finds the most capable open-weight models trail the best closed models by roughly four to seven months in cyber-related abilities.2 min read
SafetyAnthropic says Claude models are highly resistant to prompt injectionAnthropic announced that they have successfully trained their models to resist prompt-injection attacks that previously fooled early versions of Claude; similar results were observed in a benchmark…1 min read
SafetyAI models and the need to read code: bugs, technical debt and infrastructureAn AI developer comments that if you don't read code written by models, it usually indicates beginner-level, prototype, throwaway software or projects without users/revenue; this can lead to technical debt and risk.1 min read
SafetyResearcher develops printable patterns that can evade common surveillance detectorsBill Swearingen says he used millions of automated tests and a reinforcement-learning model to generate visual patterns that prevent many widely used detection algorithms from flagging people and vehicles.4 min read
