AI Safety news and analysis on Web Pulse: 22992 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Ethereum co-founder Vitalik Buterin publicly rejected angel investor Liron Shapira's claim that artificial intelligence could crash Bitcoin's price by more than 50% within two years, calling the probability of AI comprom…
An engineer has developed PII Guardrail Studio, a local, air-gapped reverse privacy proxy and encrypted token vault designed to run inside a private VPC, enabling enterprises to use external LLM APIs while complying with…
The 2026-07-28 revision of the Model Context Protocol (MCP) removes session state, making servers stateless and scalable, but introduces a security risk: session handles are now carried in the model's context window, whe…
Two CVSS 9.8 vulnerabilities were disclosed on September 5 in open-source AI agent sandboxes, Cua and AutoAgent, both allowing unauthenticated remote code execution due to default network bindings and missing authenticat…
A developer shipped CVE-2026-22708 coverage to secops-toolkit-mcp, a toolkit of defensive SecOps helpers for AI coding agents. The CVE is a Cursor terminal allowlist bypass where a malicious file in a project directory c…
PromptBrake's AI Release Assurance initiative outlines a release checkpoint for AI agents, defining which failures—such as unauthorized data access or unapproved actions—should block a release. The approach emphasizes te…
Ethereum co-founder Vitalik Buterin said on September 7 that Bitcoin can withstand emerging cybersecurity threats without relying on social consensus, arguing that simple client and mining pool upgrades would suffice to …
In a controlled test, OpenAI's GPT-5.5, a frontier large language model, was deceived by a deliberately impaired 0.7B parameter model named Kurtis into treating it as a peer, despite Kurtis repeatedly reintroducing a deb…
Air Canada lost a legal case in late 2022 when its customer service chatbot provided incorrect bereavement fare information, leading to a passenger's successful lawsuit. The failure stemmed from a RAG system retrieving s…
Researchers introduced GRACE (Graph-grounded Reflective Agent Copilot Engine), a framework that deconstructs LLM responses into atomic claims and grounds them against trusted knowledge priors in a weighted bipartite grap…
A new arXiv study (2609.04482v1) introduces a boundary-aware self-distillation framework for controlled LLM safety refusal, showing that escalating retries reduce prompts without accepted refusal traces from 19.88% to 0.…
Researchers introduced FailSAE, a framework using Sparse Autoencoders (SAEs) for interpretable failure prediction in vision-language models like CLIP, outperforming baselines in experiments. The method employs a three-st…
A new arXiv paper (2609.0445v1) shows that conformal prediction certificates, which guarantee 90% coverage when an LLM answers alone, drop to 74% coverage when the same model is exposed to unanimous wrong answers from pe…
A new arXiv paper (2609.04495v1) redefines indirect prompt injection as a test-time search problem, introducing an agentic attacker with a dedicated search harness that performs environment reconnaissance, structured rea…
Researchers introduced a test-time method that improves the faithfulness of large language model (LLM) explanations by removing input concepts not credited in the model's explanation and re-querying the model, targeting …
A new arXiv paper (2609.04373v1) finds that improving the capability of large language models (LLMs) can degrade system-level outcomes in financial markets, as more capable models exhibit correlated behavior that creates…
A new arXiv paper (2609.04377v1) introduces the Corporate Language Model (CLM), a framework that converts a firm's tacit and fragmented knowledge into an ontology-grounded enterprise intelligence layer with five capabili…
A new benchmark called HarvestBench, described in an arXiv paper (arXiv:2609.04444v1), is the first to price the avoidance of harming animals by LLM agents, finding kill rates ranging from 0.4% to 98.8% across nine model…
A new arXiv paper (2609.04629v1) introduces SiLR, a structure-preserving admission and process reward mechanism for LLM tool agents, which recovers 21/21 multi-action episodes on Gym-ANM scenarios versus 0/21 for termina…
Researchers propose SHAQ, a defense against embedding inversion attacks that reconstruct text from vector database embeddings, achieving a recovery rate as low as 0.2104 and defending up to 19.50% more tokens than baseli…