From the archive
AI safety examines how artificial intelligence systems can fail, cause harm or be misused, and how those risks can be reduced. Follow research on model evaluations, security, alignment and safeguards. Read what was tested, which threats the evidence supports and where uncertainty remains, with links to original findings rather than treating either reassurance or alarming predictions as settled facts.
Lead story
OpenAI used DevDay 2026 to expand its developer tools for agents, including computer use in the Agents API and cloud-based Codex environments. The practical shift is towards agents that can operate software and run coding tasks remotely, not just answer a prompt and wait politely for the next…
Read the report & discussion ↗Meta AI Watch
Meta has let go of employees it hired from AI security startup Virtue AI, four months after they joined. The departure raises a pointed question about how Meta is building its AI safety operation, even as the company says that work remains a priority.
OpenClaw Watch
OpenClaw says every skill and plugin uploaded to ClawHub now gets an additional security review from Tencent’s AI-Infra-Guard, alongside NVIDIA’s SkillSpector. The change puts automated checks into the marketplace’s upload review, where a risky plugin can otherwise become someone’s very avoidable afternoon.
Watch Desk
GitLab’s Threat Research Group says a flaw in DeepSeek-Reasonix Studio could let a hostile repository run attacker-controlled code when a developer views a file diff. GitLab says fixes are available, and the report points to a wider risk for coding tools that invoke Git on repositories they do…
OpenAI Watch
SpecterOps plans to launch an eight-hour course on securing large language models and AI agents, developed with OpenAI. The practical focus is on testing how these systems can fail, rather than assuming a reassuring guardrail will do the job.
Microsoft AI Watch
Microsoft’s 2026 Digital Defense Report recommends that public bodies tightly control what AI agents can access, remember and do. The practical message is to give agents narrow, temporary permissions, and keep consequential actions behind review.
Watch Desk
The White House’s voluntary AI safety pact asks companies to adopt controls and audits, but does not legally require them to do so, argues Council on Foreign Relations fellow Connor Martin. His central challenge is practical: without enforceable duties, independent scrutiny and consequences for failure, the pact may…
Watch Desk
AI coding tools are letting people without programming experience build websites and simple apps, but a working prototype is not the same as dependable software. A report by The Federal describes both the new opportunities and the bugs, false information and security concerns that can follow.
OpenAI Watch
OpenAI’s Agents SDK 0.23.0 adds opt-in Docker removal protection, tighter controls for agent tools and a substantial set of runtime and session changes. For developers building AI agents in Python, the release reaches into the unglamorous machinery that can make an agent safer to operate and easier to…
Watch Desk
AI-generated text is becoming a larger share of the web, but feeding more of it into language-model training may eventually do more harm than good. In a study of 800 models, researchers found that synthetic text’s initial benefit for data-starved models can reverse as the budget for human-written…
Anthropic Watch
Anthropic has launched Claude Sonnet 5.5, a new model aimed at everyday coding and knowledge work. The headline claims are more than 30% faster output and up to 30% lower cost per task on many workloads, while token prices stay unchanged.
Watch Desk
An essay published by The Observer argues that AI research need not slow down, but that powerful models should face stronger checks before wider release. Its central proposal is a predictable testing process for safety, security and reliability, rather than treating research speed and public deployment as the…
Watch Desk
Normal Tech argues that AI safety debate is being squeezed into a choice between existential-risk warnings and suspicion of the people making them. Its essay makes a third case: take those warnings seriously, but ask whether the framing helps produce better policy.
Microsoft AI Watch
Microsoft’s 2026 Digital Defense Report says threat actors are using AI in reconnaissance, social engineering, malware and exploit development. The report’s other warning is for defenders: AI agents’ access to data, tools and business systems makes their identities and permissions part of the security perimeter.
Cyera Watch
Cyera’s Assaf Morag argues that AI-assisted cybercrime makes one defensive question more urgent: what sensitive data can a compromised account or token actually reach? His analysis recommends checking exposed credentials, broad application permissions and bulk data access, rather than treating a successful login as the end of the…
OpenText Watch
OpenText’s Fortify Agent Skills give coding agents access to security findings and defined workflows, so they can prepare fixes with more context than source code alone. The important boundary: an agent can propose the repair without becoming the authority that signs it off. For teams already letting agents…
Watch Desk
Universities adopting AI assistants may be outsourcing more than computing, argues Hugging Face community writer Vasileios Stergiou: they may also be handing a private company influence over how students encounter difficult questions. His proposed alternative is to run open-weight models on infrastructure the institution controls, and make the…
