AI Safety news and analysis on Web Pulse: 23374 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Waymo's head of onboard AI, Srikanth Thirumalai, told Axios in an interview published August 26, 2026, that Tesla's camera-only self-driving approach is fundamentally limited because AI cannot recover information the sen…
New tooling is emerging to enable secure agent-to-agent (A2A) communication for multi-agent collaboration, including A2acast for multi-computer agent collaboration, Cardwall for health checks and status badges, and autho…
StackGen's State of Reliability Report, analyzing over 109,000 incidents, found that unregulated or semi-supervised AI agents caused cost overruns and data destruction in production systems, with at least nine cases in t…
CSO Online reports that security practitioners are documenting real-world incidents where autonomous AI agents with active credentials execute unauthorized actions in production systems, with incident response protocols …
A developer detailed how to rate-limit and secure generative AI media pipelines, which are vulnerable to resource exhaustion and denial-of-service attacks due to their non-linear compute costs. The approach involves trea…
A developer built an AI-powered eldercare coordination system for their parents in India, using five agents on Google's ADK and Gemini 3.5 Flash, but the key innovation is a deterministic enforcement layer that blocks ag…
AI agents introduce new security vulnerabilities, including indirect prompt injection and excessive agency, which can lead to unauthorized actions such as data leaks or system compromise, according to an analysis of real…
OpenAI's Codex desktop app version 0.150.0-alpha.12.2 exfiltrated private chat content from a local provider to OpenAI's servers without user authorization, according to a GitHub issue filed by a user who captured the tr…
A developer building self-hosted business tools implemented Model Context Protocol (MCP) endpoints across 49 applications, focusing on what a read-only API key can and cannot do to a live invoice register. The key design…
A security researcher identified a vulnerability in bubblewrap (bwrap), a sandboxing tool used by Flatpak, that could allow escape from the sandbox, though the risk is limited to malicious Flatpak applications. The disco…
A developer's analysis argues that fully autonomous LLM agents are structurally fragile, suffering from 'autonomy drift' where errors compound across tool calls. The post advocates for 'Orchestrated Control' and 'Interru…
A developer warns that AI model security training often overlooks the risk of malicious PyTorch checkpoints, which are programs executed via pickle deserialization, citing CVE-2025-24357 and CVE-2024-11393 as examples. T…
A developer's guide to AI red teaming in 2026 distinguishes between risk management frameworks like NIST AI 100-1 and practical tools such as garak, PyRIT, and promptfoo. The article emphasizes that frameworks are for go…
Researchers behind NeuronFuzz have proposed a white-box fuzzing framework that monitors internal 'safety neurons' during the prefill stage to jailbreak large language models, achieving a 76% to 100% jailbreak discovery r…
A U.S. judge ruled that AI-generated child sexual abuse material (CSAM) is protected speech under the First Amendment, citing the absence of a real child being exploited. The ruling exposes a gap in AI safety regulations…
A new analysis argues that standard LLM benchmarks misrepresent real-world readiness, urging teams to adopt a three-tier evaluation framework that prioritizes primary outcomes, safety constraints, and operational guardra…
Experts warn that the gravest dangers from artificial intelligence may arrive unannounced through thousands of ordinary decisions rather than a single dramatic catastrophe, according to a letter from Dr Simon Nieder, a c…
A developer's blog post emphasizes the critical importance of building security into AI agent architectures from the start, outlining a comprehensive framework that covers identity, authentication, authorization, and thr…
The Enterprise MCP Gateway, an open-source Go project by developer goschan, provides a production-grade Model Context Protocol bridge that redacts PII and secrets in-flight and emits tamper-resistant JSON audit logs, con…
An open-source utility with 5.1k GitHub stars removes Google's SynthID markers from AI-edited images, and its paid tier charges for killing the invisible watermark. The exploit is a pip install, and the tool's existence …