AI Safety news and analysis on Web Pulse: 13090 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Berkeley RDI and over 300 industry experts launched Agents' Last Exam (ALE), an open benchmark that evaluates AI agents on economically valuable professional tasks rather than abstract proxies. The benchmark currently co…
Agent frameworks like LangGraph, Google ADK, and Mastra advertise durable execution through checkpointing, but their state-snapshot approach only recovers to the last boundary, not the exact failing step, leaving develop…
Daniel Cormier's X account posted fake screenshots of Eric Trump asking if UFC fights are rigged, which Eric Trump denied as AI-generated. Cormier later deleted the post and said his account was hacked, calling the image…
Researchers have discovered that deep-research agents, which use multi-agent pipelines to retrieve and synthesize web content, can be poisoned by adversaries appending crafted text to frequently retrieved user-generated …
Anthropic pulled access to its Mythos 5 and Fable 5 AI models after the US government imposed export controls due to a jailbreak vulnerability. The company disagreed with the action, warning it could halt new frontier AI…
A developer released Pantheon, an open-source multi-agent harness for Claude Code that runs coding tasks through parallel implementations and adversarial verification to catch bugs that single-pass models miss. The tool …
New research from Cornell University shows that a single 13-word Reddit comment can manipulate AI search tools like ChatGPT and Google AI search, poisoning their outputs with promotional content. The study reveals that d…
Puppet has published guidance on remediating 18 OpenSSL vulnerabilities at scale using its infrastructure automation platform. The June 2026 advisory includes a high-severity heap use-after-free bug in PKCS7_verify() tha…
1Password acquired Apono, a just-in-time access governance startup, to govern AI access. The deal brings intent-based, zero-standing-privilege access controls into 1Password's Unified Access strategy and enhances its new…
Nick Creighton, an operator who ships, provides a detailed blueprint for deploying Meta's Llama 3 706B model locally, emphasizing privacy, latency, and cost benefits over cloud APIs. He outlines the exact hardware requir…
Ildar Iskhakov, CTO of an unnamed company, published a catalog of ten prompt injection techniques that exploit the inability of large language models to distinguish trusted instructions from untrusted data. The attacks r…
NewCore emerged from stealth with $66 million in funding to build an identity-security platform that governs both human employees and autonomous AI agents under one architecture. The Tel Aviv and San Francisco-based star…
Depthfirst's autonomous security agent discovered 21 zero-day vulnerabilities in FFmpeg, including several that had been latent for 15 to 20 years, after intensive analysis by Google and Anthropic. The agent produced con…
A developer warns that selfies, especially peace sign gestures, can expose fingerprints that AI tools can extract and replicate, enabling presentation attacks on biometric systems. Historical cases include the 2013 defea…
Equixly launched a Model Context Protocol integration that embeds its continuous penetration testing platform directly into AI coding assistants like GitHub Copilot and Claude, allowing developers to trigger security tes…
Pokémon Go players' voluntary 3D scans trained Niantic Spatial's AI models for GPS-free navigation, which are now being combined with defense contractor Vantor's software to guide military drones when satellite signals a…
NewCore, an identity security startup co-founded by Zohar Alon, Amihai Neiderman, and Erez Yarkoni, emerged from stealth on June 15 with $66 million in seed funding to secure a workforce that includes humans, machines, a…
A new AI readiness radar framework warns that AI coding tools will amplify existing engineering dysfunctions rather than fix them, urging teams to assess their foundations in codified standards, feedback loops, and trust…
Anthropic disabled its Fable 5 and Mythos 5 models worldwide on June 12 after a US government export-control order. The order, citing national security, bars access by foreign nationals, forcing Anthropic to cut off all …
A developer traces the rapid evolution of LLM interaction techniques from 'prompt engineering' to 'context engineering' to 'harness engineering,' where models autonomously iterate and learn. The post warns that as LLMs g…