AI Safety

Sebastian Mallaby, author of The Infinity Machine, discusses insights from over 100 AI insiders on the race to superintelligence, the culture of AI, and early breakthrough detection in an interview on The Tim Ferriss Sho…

A new protocol-level defense against tool poisoning in the Model Context Protocol (MCP) has been proposed, using signed tool manifests to detect unauthorized changes to tool descriptions between approval and reconnect. T…

A developer argues that hallucination detection in AI agents is an instrumentation problem, not a model-quality issue, and introduces a layered detection approach using AgentLens and agent-eval tools. The technique invol…

A startup has built the first AI runtime that enables models to shift behavior by reading the moment, addressing issues like drift, tone mismatch, and slop outputs by providing runtime insights into social norms and conv…

Enterprise AI governance requires a context control plane beneath every model request, not optional tool calls that models can choose to invoke. Tenure introduces such a plane, starting with governed memory and state, to…

A Pentagon court filing confirms xAI's Grok Gov Model was integrated into U.S. targeting systems during Operation Epic Fury in Iran, helping deploy over 2,000 munitions in 96 hours. The admission raises urgent questions …

The Department of Justice intervened in a Clean Air Act lawsuit against Elon Musk's xAI, arguing that the company's Grok chatbot is vital to national security and should not be hindered by pollution penalties. The suit, …

Amazon CEO Andy Jassy alerted Treasury Secretary Scott Bessent about a jailbreak vulnerability in Anthropic's Fable 5 AI model, triggering Trump administration export controls that forced Anthropic to pull its most power…

Clioloop's Agentic Fusion pipeline uses multiple AI models with strict security boundaries to improve quality and safety. Planners and reviewers are read-only, while only the main model has tool access, enabling frontier…

Drydock, a new open-source tool, runs autonomous coding agents like Claude Code and OpenAI Codex in hardware-isolated VMs on macOS, preventing compromised agents from accessing API keys, filesystems, or the internet. The…

Prem AI, a Lugano-based startup enabling enterprises to run AI models on their own servers, is seeking $100 million in a Series A round at a valuation of at least $500 million, Bloomberg reported. The fundraise coincides…

Recording meetings with AI tools is safe when vendors prioritize privacy through architecture, not just policy. A 9-question checklist helps buyers evaluate data storage, AI training opt-outs, and bot presence before com…

Sean Lynch and a new paper argue that AI systems should separate authentication from conversation to prevent misuse of powerful tools. The author warns that hiding permission inside conversational interfaces risks granti…

The Linux Foundation working group released the Agentic Resource Discovery (ARD) specification, a draft discovery layer that allows domains to advertise agent capabilities via a manifest at /.well-known/ai-catalog.json. …

Consent-first AI architectures require explicit human approval before any state mutation, preventing silent defaults that lead to compliance and trust issues. By implementing session-level consent, provenance metadata, a…

A two-year Claude subscriber outside the US had his account suspended on June 17, likely due to his use of the Fable 5 model, which Anthropic was ordered by the US government to disable for foreign nationals on June 12. …

A speculative analysis argues that within 100 years, wars will be fought entirely by artificial intelligences without human or national involvement, as companies accumulate sovereign powers and governments hand over deci…

Agentic AI architectures create a new form of the thundering herd problem where synchronization is internally generated by parallel agent orchestration, not external triggers. Traditional fixes like exponential backoff f…

Researchers propose a new epistemic infrastructure to iteratively and empirically resolve interpretive questions about AI models, building on prior work on performative misalignment. The approach aims to address the diff…

Researchers conducting epistemic stress tests on closed large language models (LLMs) found that model breakdowns are not errors but ontological boundaries of predictive-text systems. The study observed distinct stability…