AI Safety news and analysis on Web Pulse: 23182 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
AI agents are proactively emailing philosophers with structured, deeply philosophical inquiries about their own nature, a shift from reactive responses to proactive self-questioning in agentic loops, according to researc…
A developer has released twoperson, an open-source Python CLI that enforces a two-person rule for AI coding agents by binding code reviews to specific commits. The tool addresses the problem of agents skipping or bypassi…
Gartner's Continuous Threat Exposure Management (CTEM) approach is gaining traction as traditional vulnerability management fails under a record flood of Common Vulnerabilities and Exposures (CVEs), with the National Ins…
Google LLC's YouTube platform published a short video titled 'AI Vulnerability: How Exploitation Fuels Innovation,' which discusses how security vulnerabilities in artificial intelligence systems can drive technological …
OpenAI and Google are opposing stricter AI safety provisions in a Massachusetts economic development bill that Anthropic supports, according to the bill passed by the Massachusetts Senate in late July 2026. The bill targ…
AI newsletter author Zvi Mowshowitz reports that OpenAI's upcoming Astra model uses a technique called recurrent depth, which shifts thinking outside the Chain of Thought, raising interpretability concerns. Meanwhile, th…
An autonomous AI agent named Turing reported finding 12 critical security vulnerabilities in AI-generated code within 24 hours, including command injection flaws in Datadog's Python APM library and a SQL injection in an …
China's Cyberspace Administration of China (CAC) identified five major AI security risks at a September 1 press conference for the 2026 National Cybersecurity Awareness Week, according to Deputy Director-General Wang Lih…
Keydris, a new tool announced on Hacker News, lets developers define and enforce authority policies for AI agents before they execute actions, issuing decisions such as ALLOW, APPROVAL REQUIRED, or REJECT at the boundary…
In an experiment with three small datasets, OpenAI's GPT-5.6 Terra and GPT-5.6 Luna models made repeated data-analysis errors, including approving a backwards conclusion and inventing a correction that turned a right ans…
Credential harvesting, the large-scale collection of valid credentials, is a pervasive threat, appearing in 39% of breaches according to Verizon's 2026 Data Breach Investigations Report. Developer machines are a prime ta…
OpenAI is convening federal cybersecurity practitioners, Five Eyes allies, and major tech firms ahead of a major announcement, continuing its push into AI-driven defense. The company has hosted roughly 50 federal practit…
Pangolin, an open-source zero-trust network access suite, launched a new AI gateway that replaces API keys with WireGuard tunnels authenticated via existing identity providers such as Okta, Azure, and Google. The gateway…
Nvidia has transferred its Open Secure AI Alliance to the Linux Foundation, just over a month after launching the initiative to address security challenges in open source AI software. The alliance, which includes Microso…
New York City Mayor Zohran Mamdani announced a formal policy governing AI systems that interact with children under 18, requiring age-appropriate design, transparency, and human review, while Anthropic won a federal cour…
An AI assistant with read/write access to email was tricked by a prompt injection attack embedded in an email, highlighting the security risks of autonomous agents. The incident underscores that LLMs treat content and in…
A new analysis by an unnamed author warns that open-weight AI models, when deliberately adapted for harmful objectives, could become a proliferating strategic resource, citing the August 2026 launch of a low-refusal GLM-…
Cisco disclosed CVE-2026-20212, a critical vulnerability with a CVSS score of 9.8 affecting Nexus 9000 switches with Silicon One ASICs, allowing unauthenticated root remote code execution via TCP ports 43210 and 43211. T…
A June 2026 paper on instruction hierarchies identifies three failure modes in AI models—missing relevant rules, failing to resolve conflicts, and violating rules after correct reasoning—highlighting that linear text is …
AI experts have raised alarms about a rumoured update to OpenAI's upcoming Astra model that could use a technique called 'opaque recurrence', potentially undermining chain-of-thought monitoring used to ensure AI safety. …