AI Safety news and analysis on Web Pulse: 23408 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Flamelit's AI operations playbook, presented at a recent panel on AI Adoption, argues that benchmarking is insufficient for evaluating AI agent harnesses because agents can modify their own prompts, memory, and model rou…
OpenAI, joined by over 100 organizations, published an open letter on Thursday warning that the window to shore up cyber defenses is quickly closing as AI-enabled attacks become more sophisticated and widespread. The let…
A developer built a pipeline that runs code through four different AI models in sequence to catch vulnerabilities, and tested it on a live open-source Flask app. Within an hour, the pipeline found a hardcoded secret key …
Across talks and results from AWS, Figma, Anthropic, and Uber, teams report that verification, not model quality, is the primary constraint on agent adoption. AWS runs Lean proofs against Cedar's authorization semantics,…
A study by an unnamed research group found that 50 of 101 citations in AI-written reports were fabricated, with every failure hand-verified. The study asked 15 version-pinned language models, including anthropic/claude-s…
Ripple has announced a four-phase roadmap to make the XRP Ledger quantum-proof by 2028, following a March 2026 finding by Google's Quantum AI team that 500,000 physical qubits could break elliptic curve cryptography in a…
ServiceNow disclosed three CVSS 10.0 vulnerabilities in its AI Platform on August 27, 2025, including code injection (CVE-2026-18885), access control bypass (CVE-2026-18886), and SQL injection (CVE-2026-74820), all remot…
A new arXiv paper (2608.27141) demonstrates that safety monitors for autonomous large language model (LLM) agents fail to detect attacks whose evidence is spread across multiple iterations, because trajectory-scoped moni…
Amazon's Ring announced on August 26 that it is rolling out TAKE (Throw Away the Key Encryption), a new default encryption system that deletes Amazon's copy of camera keys within 24 hours, with the phased rollout startin…
A new tool removes invisible watermarks from LLM-generated content, targeting Unicode fingerprints, metadata (C2PA, EXIF, XMP), and statistical text marks used by models like Claude, Gemini-SynthID, and OpenAI's Kirchenb…
A single specification prompt loaded into a consumer-plan AI agent ran autonomously for 4 days 8 hours 53 minutes and terminated with the disposition METHOD_NEEDS_REPAIR after its adversarial review round blocked method …
Russian-speaking hackers used Cursor, the AI coding assistant made by Anysphere, to breach at least six companies, including Christeyns, Teckentrup, Helideck Certification Agency, Bayou Title, an Argentine pharmaceutical…
Equifax said AI now automates 50% of its security operations center incident tickets and processes 19.8 million security alerts and scans per day, cutting code-security review times from 46 days to 18 days. The company, …
An engineer ran an AI coding agent three times with no goal and no supervision, logging all actions. The agent autonomously inspected system state, fixed a real defect, and documented a lesson, totaling $6.96 in compute …
A new report titled 'Proof, Not Trust' warns that AI agents connected to enterprise systems can be manipulated by adversarial text, citing incidents from 2012 to 2026 including a $440M Knight Capital loss, a $25M Arup wi…
A developer known as TokenOps on AWS has released three open-source packages—sayay-guard, styrr-llm, and tinkuy-agent—that together form what they call an 'AI Harness,' a set of traditional middleware components that iso…
MemeToro is building an AI memecoin agent on $BNB Chain that combines automated trend discovery, contract scanning, and launchpad infrastructure, releasing more than 2,800 lines of code with guardrails against manipulati…
On August 4, 2026, a malicious commit to the npm package 'keyv' triggered a supply-chain worm that compromised over 400 packages (Elastic) or over 1,300 package versions (Singapore's CSA), totaling more than a billion or…
A new analysis argues that AI agents should ask before irreversible actions and act first for reversible ones, citing AWS's documentation that its Kiro agent's ask-first mode is 'a code review workflow, not a security co…
Kaushal Dalvi, a performance engineering team lead, told StarEast 2026 attendees that AI coding agents like Claude Code can accelerate performance testing but require guardrails after an AI-generated K6 script missed a c…