AI Safety news and analysis on Web Pulse: 23562 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
According to an analysis of large language model (LLM) tool failures, only three root causes exist—value, condition, and intent—and the primary fix is to shift from validation to asking questions using an external list. …
A developer spent six weeks letting a multi-agent system run the operational backbone of a SaaS, using LangGraph for state management and a RAG system fed by Jira, GitHub, and Stripe data. The experiment revealed subtle …
OpenAI reportedly paused internal development of its Astra model after evaluations found it could autonomously discover and weaponize zero-day exploits, marking the first model to trigger the 'Critical' cybersecurity thr…
Nearly half of organizations that have deployed AI-based security controls still experienced confirmed or suspected AI-related incidents, according to the 2026 Proofpoint AI and Human Risk Landscape Report, signaling tha…
The Cybersecurity and Infrastructure Security Agency (CISA) added CVE-2025-62593, a critical remote code execution flaw in the Ray AI framework, to its Known Exploited Vulnerabilities catalog on August 17, 2026, after co…
Linus Torvalds detailed a debugging session where Google's Gemini repeatedly claimed a Linux kernel bug was impossible to fix, yet he persisted and found a one-character fix, now merged into Linux 7.3. Separately, physic…
Snowflake released CoCo Automations into public preview on August 21, allowing scheduled, unattended AI agent runs, but the default grant of EXECUTE AGENT TASK to the PUBLIC role exposes all users to potential safety ris…
A new arXiv study evaluating therapy chatbots found that large language models from Claude, GPT-4o, and Llama-3.1 understand 76-82% of Generation Alpha mental health vocabulary but correctly calibrate only 64-72% of clin…
Researchers introduced a data-poisoning attack called cross-parameter relinking that creates wrong-physics backdoors in neural PDE operators, making triggered inputs output physically plausible but incorrect solutions. I…
Researchers introduced aiXamine, a unified black-box platform that evaluates LLM trustworthiness across safety, security, and privacy, and found that stronger safety alignment increases over-refusal, privacy is near-orth…
Researchers introduced Thermo-FL, a thermal-aware federated fine-tuning framework for large language models on edge devices, which uses device temperature to control local adapter training and sparse update transmission,…
Zabriskie, a social app built through vibe coding, discovered that its home screen scheduling algorithm failed to show 20 of 47 applicable cards to a user across five daily programs, despite passing implementation, tests…
Researchers propose TH-GNN, a heterogeneous temporal graph neural network that detects LLM-agent shilling attacks on recommender systems by jointly modeling temporal, structural, and semantic signals. In evaluations acro…
Hackers are actively exploiting CVE-2026-64849, a critical unauthenticated Server-Side Request Forgery (SSRF) vulnerability in MLflow, an open-source AI development platform used by thousands of organizations including M…
Kaspersky's Global Research and Analysis Team (GReAT) warned that blind trust in AI agents is fueling a new wave of supply chain attacks, with 92,000 malicious attacks in 2026 disguised as legitimate AI services, 49% imp…
Security researchers warn that fake AI tools disguised as productivity apps, coding assistants, or browser extensions are a leading malware vector in 2026, with attackers exploiting the fast-moving AI market to trick use…
In March 2026, threat actor group TeamPCP published two backdoored versions of LiteLLM (1.82.7 and 1.82.8) to PyPI, live for roughly six hours, compromising more than 2,500 organizations and 434,000 CI/CD pipeline runs, …
OpenAI Group PBC Chief Executive Sam Altman warned in a podcast interview on Sunday that artificial intelligence could end up controlled by a handful of companies or people, leaving society without a voice in how it impa…
Xingyao Byte, a developer-focused firm, argues that prompt injection cannot be solved at the prompt layer due to the architectural nature of LLMs, and instead advocates for system-level defenses such as separating privil…
WebAZ, a commerce platform, has introduced a dual-interface architecture where a Progressive Web App (PWA) serves human users and a Model Context Protocol (MCP) server serves AI agents, both operating over a single state…