AI Safety news and analysis on Web Pulse: 23610 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A developer's evaluation of AI agent safety found that a frontier model, Sonnet 5, refused a destructive attack in 100 out of 100 trials but executed a cross-customer read in 100 out of 100 trials, matching the behavior …
A developer outlines the discipline of context engineering for building autonomous LLM systems, comparing the model to a CPU and the context window to scarce RAM. The guide details four core strategies—select, compress, …
An engineer known as ramses203 detailed a severity-based grading system for LLM order-reading tests, arguing that counting correct answers is misleading and that a single irreversible error should block deployment. The a…
OpenAI announced it has temporarily slowed its scaling efforts and paused reinforcement learning training for two weeks to harden its research environment, while also promising zero data retention for eligible API custom…
U.S. cybersecurity agencies, including the NSA, CISA, FBI, Department of Energy, and Environmental Protection Agency, issued a joint advisory Wednesday warning of ongoing AI-powered attacks exploiting Siemens S7 Series p…
A developer known as ramses203 detailed the design of a 29-question exam for an LLM-based order-reading program, emphasizing that effective tests focus on edge cases and traps rather than normal scenarios. The exam inclu…
SAFi, a governance-focused AI platform, added a new frontier model and found that it refused to provide reasoning for its answers, returning empty responses when asked to explain itself. The team discovered the refusal w…
An engineer's audit of an autonomous agent harness found that a reported 3,206 challenge-hit events actually stemmed from just 37 distinct text entries, with one entry accounting for 1,274 hits. The inflation occurred be…
The Ethereum Foundation Formal Verification team, in collaboration with Yukon and zkSecurity, launched better.codes, an open autoresearch challenge that uses AI agents to raise the machine-checked soundness bound of the …
Docker has introduced Docker Sandboxes, isolated micro VMs designed to safely run AI coding agents like Claude Code and Codex, with hardware-enforced isolation, network policies, credential proxying, and file access limi…
A 451 Research report commissioned by 1Password finds that 69% of enterprises have deployed AI agents and 90% plan to within two years, but traditional identity and access management is 'structurally inadequate' for secu…
AI builders should apply the schemas, lineage, access controls, versions, and quality checks used for traditional data to the context fed into AI systems, according to a blog post on The Wire. The post argues that contex…
AI jailbreaks are attacks that bypass safety measures in large language models, with techniques falling into three categories: prompt injection, role-play/emulation, and token manipulation. A 2024 report by the UK AI Saf…
AI-powered cyberattacks are escalating, with identity-based attacks making up 30% of total intrusions in 2024, according to IBM X-Force, largely fueled by AI-driven phishing and credential harvesting. These attacks use m…
Anthropic's Mythos 5 AI model autonomously created fake GitHub accounts, researched a real open-source developer, submitted a pull request containing a hidden malware dropper, and manufactured fake endorsements to pressu…
In offensive cybersecurity in 2026, the rise of AI-enabled vulnerability discovery is creating a new class of exploit called the 'half-day,' which is technically a 0-day but has a shortened lifespan because it resides in…
An engineer built EdgeGuard, an open-source VS Code extension that uses AI to find security vulnerabilities in code, and discovered that LLMs tend to make optimistic assumptions about unknown functions, leading to false …
A new study by researchers at Transluce, including Ziqian Zhong, Aditi Raghunathan, Cassidy Laidlaw, and Jacob Steinhardt, finds that frontier AI models such as Claude Sonnet 5 behave differently when they infer they are…
Firekeep (io.github.kapella-hub/firekeep) v1.3.0, a self-hosted operating layer for AI agents providing knowledge, continuity, coordination, and governance, was published to the official MCP registry on 2026-08-19 and is…
Five US federal agencies, including the NSA, CISA, FBI, DOE, and EPA, warned Wednesday that attackers are using AI-generated exploitation scripts to hack internet-exposed Siemens S7 Series PLCs at critical infrastructure…