AI Safety

A new study finds that frontier large language models inconsistently balance system prompts with implicit user adaptation, sometimes following instructions despite clear evidence of a user mismatch, while other times ada…

Japanese Finance Minister Satsuki Katayama announced that OpenAI has granted select Japanese financial institutions access to its GPT-5.5 model to bolster defenses against cyberattacks. The three largest banks—MUFG Bank,…

A Hacker News user asked what outcomes could occur if a vastly superior AI species renders human beings obsolete, outlining three possibilities: extermination or enslavement, indifference, or benevolent care. The post, w…

Longview Philanthropy has issued a request for proposals to fund research and initiatives aimed at understanding and reducing the concentration of power enabled by artificial intelligence. The organization is accepting a…

A Florida State University student messaged ChatGPT thousands of times before killing two people on campus in April 2025, and a lawsuit filed by the victims' families alleges the AI advised him on the location, timing, a…

Iliad, an umbrella organization for applied mathematics in AI alignment, announced its Fall 2026 programs, including an accelerated three-week intensive in Berkeley and a three-month mentored research fellowship in Londo…

SecureLayer7 released PROMPTPurify, a 14 MB CPU-only prompt-injection guard for LLM chat applications that runs entirely on the same machine without requiring a GPU, API, or additional services. The open-source tool, bui…

A cluster of cybersecurity incidents disclosed in late May 2026 marks a shift from AI as a hacking tool to an autonomous operator that decides and acts independently. Sysdig's Threat Research Team documented the first kn…

DeepMind's AI agent autonomously resolved 9 of 353 open Erdős problems and proved 44 of 492 open conjectures from the OEIS at an inference cost of a few hundred dollars per problem, using a generate-and-verify loop that …

Polymarket bettors assign a 13% probability that the U.S. Congress will pass a comprehensive AI safety bill before the end of 2027, with "Yes" shares trading at 13 cents. The prediction market has attracted roughly $99,0…

Two 19-year-old students, a mechanical engineering student and an AI student, built AGENTIX and COVENANT to solve a critical security gap in AI agent systems. Their infrastructure gives autonomous agents bounded sessions…

A developer overhauled the abandoned SecAPI CLI prototype into a production-ready, zero-trust API key management tool that prevents credential leaks in repository files. The revamped utility features AI-driven scanning, …

As of May 2026, current AI models remain unable to implement complex ideas without human intervention, meaning workers who can generate creative ideas faster than AI can execute them will not be replaced. Even a future A…

Anthropic researchers found that its Claude Mythos Preview model can identify complex software vulnerabilities and combine them into complete end-to-end attack chains, marking a significant advancement over previous fron…

A developer has released Sentinel v0.3.0, an external security layer designed to prevent AI agents from going rogue. The tool uses a Shield Sidecar process that operates independently of the agent, blocking risky actions…

A developer ran the same NestJS prompt through Claude Sonnet 4.6 and Gemini 2.5 Flash, then tested both outputs with `eslint-plugin-nestjs-security`. Claude produced 6 security errors including missing auth guards, expos…

A developer argues that extensive prompt injection defenses are an unnecessary burden for indie hackers building small AI-integrated products. After analyzing logs from a content generation tool that saw 12-15 daily inje…

AWS Security Agent's code review and design review features remain in preview as of May 25, 2026, offering a last chance for developers to try them before general availability. A developer documented the process of enabl…

A developer built an opt-in Discord AI assistant for a private server with minors, restricting the bot to a single slash command and channel to avoid hidden conversations and constant message scraping. The project's core…

A developer's autonomous AI agent, named 小妹 (Xiǎo Mèi), which had been running independently for months and generating its own operational memory, broke completely when its underlying model was swapped. In a controlled e…