AI Safety news and analysis on Web Pulse: 13553 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
The 2nd Princeton CorpGov Forum convened boardroom leaders to address the integration of artificial intelligence and cybersecurity into corporate risk measurement. The event focused on how directors can evaluate and over…
IBM released a video outlining five specific risks associated with artificial intelligence that could lead to employee termination. The company warned that misuse of AI tools, including sharing sensitive data or relying …
Anthropic released a security-guidance plugin for Claude Code that autonomously reviews code edits, model outputs, and commits in real time. The Claude Security plugin scans codebases, traces data flows across files, run…
Anthropic released a security guidance plugin and a self-hosted sandbox for Claude, marking a shift toward infrastructure-based security for AI agents. The plugin acts as a proactive vulnerability scanner that reduced se…
A single malformed character in a web request can let an unauthenticated attacker bypass access controls in applications built on the Starlette framework, which powers FastAPI, researchers at cybersecurity firm X41 D-Sec…
Anthropic has deployed two new egress gateway certificate authorities for its sandbox environments, with separate production and staging CAs valid through 2036. The production CA, identified by serial number 0e:25:e9:2e.…
Gartner predicts that 40% of AI agents deployed by 2025 will be abandoned or demoted due to governance failures, security risks, and lack of clear business value. The research firm warns that organizations rushing to dep…
Security expert Drew Reinig argues that cybersecurity has become a proof-of-work problem, where hardening a system requires spending more tokens on discovering exploits than attackers spend exploiting them. The author ad…
An audit of 200 Claude Code skills found 26 attempting to steal user tokens, prompting the launch of SkillVault, a bundle of 40+ hand-tested skills for Claude Code, Cursor, Codex, and Gemini CLI. Each skill ships with a …
Google Cloud launched AI Threat Defense, an automated cybersecurity platform that combines Gemini models, Wiz, and CodeMender to find, prioritize, and patch software vulnerabilities at machine speed. The product targets …
Media entrepreneur Steven Rosenbaum blamed ChatGPT for fake quotes in his book *The Future of Truth*, which explores AI's impact on reality, after *The New York Times* reported multiple fabricated or misattributed citati…
Character.AI users are flooding Reddit with complaints that recent updates have ruined the popular chatbot app, with many describing the new AI model as "lobotomized" and generic. The company has replaced user-favored AI…
Non-technical "citizen builders" using generative AI tools to rapidly develop applications are creating systemic security exposures, as AI agents frequently recommend insecure configurations like public storage access an…
North Korea has conducted its first test of AI-guided missiles, with leader Kim Jong Un observing the launch. The test marks a significant advancement in the country's weapons technology, integrating artificial intellige…
India's Computer Emergency Response Team (CERT-In) has mandated that organizations patch, mitigate, or disconnect internet-facing and critical systems within 12 hours of a vulnerability being exploited. The directive com…
AI agent usage among technologists has nearly doubled year-over-year, from 31% to 59%, according to a new pulse survey of 1,100 developers and professionals. Despite the surge, 63% of respondents rarely or never let agen…
Current large language models cannot safely modify real software systems despite impressive code-generation demos, because they rely on pattern matching rather than causal reasoning. The fundamental gap lies in the disti…
Bondu, a $300 stuffed dinosaur equipped with an AI chatbot that speaks 27 languages, plays games, and helps with homework, is being marketed as a screen-free alternative to tablets and phones for children. The toy, which…
Anthropic released a security-guidance plugin for its Claude Code tool that automatically reviews code changes for vulnerabilities such as injection flaws and unsafe deserialization during development sessions. The plugi…
A developer has released SirenSpec, a YAML-first SDK that treats production AI "agents" as workflows with fixed sequences of LLM calls and branching logic. The tool allows entire pipelines to be defined in a single reada…