AI Safety news and analysis on Web Pulse: 28059 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Render launched Render Sandboxes in public beta, giving AI agents a fully isolated Linux environment built on Firecracker microVMs where they can install dependencies, edit files and run code, with each sandbox starting …
OpenAI reported that an autonomous vulnerability-discovery pipeline built by its security researchers validated and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel in a single run…
A developer released Relaymote, an MIT-licensed open-source companion daemon for Claude Code sessions, along with a 38-line request guard that blocks browser-based attacks against its local control API on 127.0.0.1. The …
The RigorMed team published a self-written synthetic teaching example showing how a small standard-library Python worksheet can compute named differences between two group summaries without asserting a causal effect. In …
AWS has open-sourced Strands Box, a local AI agent sandbox that combines kernel-level OS isolation with Dogwood temporal policies, which block agent actions based on session history. The release gives developers a tool f…
DNSFilter's "The Cost of Shadow AI" report found that 74.5% of respondents have a larger cybersecurity budget than 12 months earlier, and 40.3% said their next incremental security dollar would go to AI security and gove…
A controlled study of 10 LLM models across 1,000 synthetic search worlds found that final answers are more sensitive to the order of initial evidence during active search than when all evidence is supplied in full, accor…
A new arXiv paper (2611.05814v1) introduces Consultation Value Distillation (CVD), a method that estimates offline, via Monte Carlo executor rollouts, how much an LLM agent benefits from consulting a human at a given dec…
A comparative evaluation of six large language models found authorship impersonation remained challenging for most, with Gemini 3 Pro showing the strongest apparent mimicry but only because it extensively copied and reus…
An evidence gate that lets a reviewer model override an upstream agent only after producing a mechanically valid record of a binding-constraint violation and a compliant repair sharply reduced corruption of correct propo…
A pre-registered arXiv study (2611.05577v1) of 2D decaying Navier-Stokes flow found that learned fluid representations can contain physical quantities without making them controllable, with a JEPA world model failing its…
A new arXiv paper (2611.05524v1) introduces MA-SafeDiffuser, a multi-agent extension of SafeDiffuser that gives diffusion-based trajectory planners finite-time diffusion invariance guarantees for joint-agent safety speci…
A new arXiv paper (2611.05586v1) proposes three Sybil-resistant rating algorithms for agentic marketplaces, deriving provider ratings from existing onchain payment histories rather than human curation, identity credentia…
A new arXiv paper (2611.05632v1) presents Assurance-Conditioned Continuous Authorization (ACCA), a harness-independent approach for maintaining operational authority over autonomous AI agents based on currently justified…
A study posted as arXiv:2611.05609v1 finds that multi-teacher on-policy distillation (MOPD) and on-policy distillation (OPD) can consolidate in-distribution performance while leaving out-of-distribution generalization be…
A study posted as arXiv:2611.06116v1 found that the harness around a fixed open-weight language model can swing its compliance with banking conduct rules by as much as 11.5 percentage points, enough to move one model fro…
A study posted as arXiv:2611.06545v1 found that LLM agents' self-written handoff summaries sometimes assert permissions or requirements the underlying record does not support, and that neutralizing false approvals cut un…
A new arXiv paper (2611.06806v1) proposes detecting observable AI behaviors that act as risk factors for psychological harm in LLM conversations rather than trying to measure harm directly, operationalizing eight such be…
A new benchmark, TASC (Temporal Action Safety in Conversation), finds that across 13 large language model assistants, models account for time-dependent constraints in only 30.5% of cases where a proposed action violates …
Joe Security, the Swiss malware-analysis firm founded in 2011, launched Adaptive Detonation, a feature in its Joe Sandbox AI plugin for Joe Sandbox Cloud that uses AI agents to identify why a sample stays inactive and ap…