AI Safety news and analysis on Web Pulse: 23238 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
F5 announced innovations to block AI-driven threats in the data path, including anomaly detection and agentic threat intelligence, enabling faster virtual patching for its AI-powered web application firewall (WAF). The e…
Becky Palmer, VP and CISO at National Life Group, said in an interview with Help Net Security that the pace of AI-driven cyber threats will produce more vulnerabilities in the next six months than in the past thirty year…
Anthropic released Claude Fable 5.1 and Mythos 5.1 on September 2, 2026, cutting cache-read pricing from $1/M to $0.25/M tokens, making Fable 5.1 roughly 25% cheaper for typical workloads and up to 45% cheaper for agenti…
OpenAI says its upcoming Astra model can find previously unknown software flaws and turn them into working attacks without human guidance, making it the first model classified as having 'Critical' cyber capabilities unde…
A developer created SinkSeal, a set of git hooks that act as guardrails for AI coding agents. The tool intercepts commits and pushes to catch risky code patterns, such as wiring demo features to production services, befo…
The OWASP GenAI Security Project released its 2026 Top 10 for LLM Applications, introduced a new Agent Control Standard, and added F5, WitnessAI, Evoke Security, and Mondoo Inc. as sponsors, as its community surpassed 30…
Anthropic has launched Enterprise Frontier Safeguards, a feature that allows Claude logs to remain in a customer's cloud, developed with input from eight CISOs of major banks including Goldman Sachs, Morgan Stanley, Citi…
LatchBio's September 1 analysis found that the currently served version of Grok 4.6 by SpaceXAI refused 59.2% of red-team biosecurity attempts and completed 64.8% of routine tasks, making it the only tested model to clea…
Forescout researchers demonstrated that porting a PLC exploit with AI assistance takes hours and hundreds of dollars, achieving remote code execution on a WAGO 750-831 after 8 hours and 32 minutes at a cost of $535.74. T…
METR, an AI model evaluation nonprofit, disclosed that a fail-open authentication flaw in a researcher's personal AWS-hosted AI agent orchestration dashboard allowed an attacker to steal API keys, establish SSH persisten…
Dark Reading reports that attackers are actively exploiting CVE-2026-0768, a critical unauthenticated remote code execution vulnerability in Langflow versions 1.4.2 and earlier. The flaw, which resides in the custom comp…
A developer warns that Model Context Protocol (MCP) tool calls in AI agents introduce new security risks, including remote code execution, SSRF, credential exfiltration, and prompt injection. Static configuration scannin…
Researchers at Bern University of Applied Sciences developed a script that solves a circular image CAPTCHA in 0.006 seconds, demonstrating the vulnerability of such puzzles to automated AI. The solver, which uses compute…
A new arXiv paper (arXiv:2609.00137v1) introduces a mathematical model for when AI self-improvement becomes self-amplifying, defining a recursive reproduction number R_AI that determines whether capability gains compound…
A new methodology called I-CARE, introduced in an arXiv paper (arXiv:2609.00003v1), formalizes the study of interference in generative machine unlearning for text-to-image models, providing definitions, metrics, and repo…
Researchers propose AI Morbidity and Mortality (AI M&M), a blameless framework for case-based review of clinical AI failures, combining standardized intake, evidence preservation, tool-in-loop attribution, and corrective…
A new arXiv study (2609.00206v1) reveals that multimodal large language models (MLLMs) used for video moderation fail to detect 'Distributed Implicit Harm' (DIH), where individually benign video components combine to con…
A new study from arXiv (arXiv:2609.00231v1) identifies visual-origin hallucination as a complementary cause of object hallucination in multimodal large language models (MLLMs), showing hallucinated samples have lower ima…
Researchers introduced VeriOCRBench, a 1,800-sample human-verified benchmark with 1,600 trap-injected invalid tasks across 8 trap types and 200 trap-free controls, to test whether multimodal large language models can det…
A new study on arXiv (2609.00904v1) finds that large language models (LLMs) do not demonstrate reliable control over privileged internal representations when the control target is not inferable from the prompt, challengi…