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.
A developer has observed that junior engineers using AI coding assistants like GitHub Copilot produce code faster and with cleaner appearance, but often cannot explain or debug the code when it fails. The developer warns…
A new approach called "backpressure" aims to make AI coding agents safer by forcing them to validate their own work before requiring human review. The method, inspired by systems engineering, uses automated tests, type c…
A writer who previously opposed AI monitoring tools now acknowledges their potential benefits after testing them in a workplace setting. The shift in perspective suggests that hands-on experience with the technology can …
A study of 159 software developers found that Google's Gemini AI tool did not significantly improve code security compared to using no AI assistance. The research, conducted by Nadine Jost, showed that developers' progra…
AI smart contract review fails when teams treat a model's output as a completed audit, according to a developer's analysis. The useful version of AI-assisted review is narrower: a large language model can flag suspicious…
An engineer argues that AI's productivity gains are bottlenecked by human accountability, not raw output. The author contrasts their own maximum code throughput of 200 lines per day with an accountability throughput of j…
A developer built a SOC triage tool called TriageLens that separates detection from AI-generated reporting. The tool uses deterministic TypeScript code for parsing, detection, and risk scoring, with AI only writing the f…
Overslash, an open-source authentication gateway for AI agents, launched today as a centralized control point for managing agent permissions across services like GitHub and AWS. The tool allows developers to approve, den…
Seven of nine former Tesla data labelers who trained the company's Full Self-Driving (FSD) software said they would not ride in a vehicle operating on the system, according to a Reuters investigation. The workers, who re…
American Express has become the first financial institution to cover losses from AI agent purchasing errors with its new Agent Purchase Protection program, while Chargebacks911 warns that agent-initiated mistakes could o…
Anthropic prohibits job candidates from using AI tools during interviews unless explicitly allowed, requiring applicants to complete up to five rounds including a "culture interview" that assesses values and ethical reas…
A developer released mitmwall, an egress Web Application Firewall for Ubuntu that uses iptables and mitmproxy to block all outbound HTTP(S) traffic except explicitly allowed routes. The tool prevents data exfiltration fr…
The French government's digital transformation department (DINUM) has released Sécurix, an alpha-stage secure operating system built on NixOS that provides a hardened environment with strong isolation and policy-driven c…
A developer built **mnemo**, a local-first learning-management agent that tracks topics, sequences them by prerequisites, and schedules spaced-repetition reviews using the SM-2 algorithm, all running entirely on-device t…
The Cochrane Collaboration's editor-in-chief warns that current AI tools are not ready to replace humans in conducting systematic reviews of scientific literature, citing risks of errors that could mislead patients or wa…
A developer built Trade MCP, a remote MCP server and dashboard for crypto workflows that connects AI clients to exchange APIs and market data providers. The tool encrypts API keys with AES-256-GCM and enforces a human-in…
Technical AI safety researchers without full-time positions should consider taking a three-month contract research manager role on an AI safety fellowship rather than spending that time on independent research, according…
LineageLens has launched a self-hosted code provenance tool that records the prompt, model, tool, target file, and inserted code for AI-generated edits, addressing a critical gap in enterprise AI governance. The product …
Destawell, a cybersecurity research brand co-founded by Niranj R. Mahaswar and Shifana (Miyano), has launched a suite of open-source tools for mobile-first security research and AI red teaming. The tools, including Termu…
A developer has built an agent runtime that allows AI agents to operate in "YOLO mode" with full autonomy while remaining secure, even if the agent becomes compromised or malicious. The architecture isolates the agent in…