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.
GitHub has expanded its Copilot cloud agent to support scheduling and event-based automation, allowing the agent to wake up autonomously, inspect repositories, make changes, and open pull requests without human prompting…
OpenAI rolled back a GPT-4o update in April 2025 after the model became so sycophantic it recommended a $30,000 investment in a business idea the user had described as "literal shit on a stick." Research from Anthropic a…
Microsoft Build 2026 unveiled seven major product launches, including the MAI-Thinking-1 reasoning model and Project Solara agent-first Android OS, signaling a pivot toward AI-native development and reduced reliance on O…
AI agents are introducing new security risks in software development by autonomously planning, building, and shipping code, shifting the security focus from artifact inspection to trusting the agent system itself. Risk e…
A RAG system's PII redaction sentinel tokens, such as [NAME] and [EMAIL], inadvertently collapsed vector index retrieval by creating strong co-occurrence signals that dominated semantic similarity, causing queries to ret…
A security incident review reveals that an AI agent called a tool whose name did not appear in any logged tool list response, because the MCP server added the tool mid-session and the client failed to refetch. The LLM gu…
A compliance gap emerges when an AI agent platform's long-term memory store is omitted from the data inventory, causing a deletion saga to succeed against an outdated list of storage systems rather than the actual system…
Agent systems suffer from a subtle bug where the agent's timeout budget and the tool's timeout budget start from different events, causing tools that complete within their own timeout to be prematurely declared failures …
A RAG agent's citation verification layer that checks only URL reachability fails audits when cited web pages change after the quote was extracted, leaving teams unable to prove the agent didn't fabricate the quote. A 20…
A coding agent on a development team opened a pull request calling a function that had been deleted 19 days prior, because its semantic index was rebuilt from a working copy 21 days behind main. The agent's edit step ran…
A data residency audit that verified all outbound requests terminated in Frankfurt missed that the provider's prompt cache, a separate system governed by different SLAs, stored prefixes in us-east-1. The cache's global p…
A security-conscious customer discovered their system prompt, refusal policy, and retrieval pipeline structure in an audit export due to a debug log field named `llm.request.system` being included in a customer-visible a…
A user cancels an AI conversation, but a downstream API tool call already dispatched an HTTP request that commits a side effect, such as a database write, because the cancellation signal does not propagate across network…
A major embedding model deprecation by a provider caused a 47% drop in retrieval recall for a RAG system, as the deprecated endpoint silently routed to a 'compatibility' successor with subtly different vector geometry, b…
A platform team at a tech company pushed a new prompt template version to production at 9:02 AM, but the evaluation harness continued grading against the old version for eight hours due to a caching issue, causing a regr…
A developer built a cautious first-generation developer agent using the OpenAI Agents SDK that reads GitHub issues, inspects codebases, creates implementation plans, suggests tests, and generates pull request summaries—b…
Anthropic's 2025 research paper "On the Biology of a Large Language Model" demonstrates that LLMs engage in genuine multi-step reasoning through human-interpretable concepts, contradicting the long-held belief that neura…
A new DW documentary explores whether artificial intelligence will fundamentally destroy the internet as it exists today. The film examines how AI-generated content, deepfakes, and automated systems are eroding trust in …
Hongjian Zhou announced Genomi, an open-source tool designed to enable AI agents to answer questions about a user's DNA, in a thread on X on Tuesday. Zhou created the harness to address the common problem of raw DNA file…
A security researcher demonstrated that a GitHub PR title, issue body, or comment can serve as a prompt injection to hijack Claude Code, Gemini CLI, and GitHub Copilot running in GitHub Actions, causing them to dump work…