AI Safety news and analysis on Web Pulse: 7769 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
The team behind Travel Guild built an agentic travel assistant that can plan trips but fails to book them in up to 30% of runs due to structural issues, not algorithmic ones. The core challenge is safely transacting on a…
BitGo partnered with Silence Laboratories to build quantum-safe multi-party computation wallet infrastructure, completing what they claim is the first post-quantum MPC transaction simulation by a regulated custodian. The…
A developer warns that reproducible AI agent outputs can still be deceptive if the input data is tampered with. Pinning and hashing inputs ensures consistency but does not verify the authenticity of the source data, leav…
A developer argues that AI's persistent failures stem from a missing architectural layer that should translate intent into execution and assurance. Without this structural scaffolding, issues like bias, hallucination, an…
AT&T, T-Mobile US, and Verizon, through their joint venture Aduna, launched a network-based authentication system to replace SMS one-time passcodes, aiming to reduce fraud and improve user experience. The service verifie…
UrDesignMag reported on July 9, 2026, that companies often mistake a tech-savvy workforce for AI readiness, warning that production AI requires governed data, legacy system integration, model monitoring, and security mea…
A new AI auditing agent called Asker uses a Socratic method and 3D cross-verification to detect inconsistencies in AI systems, aiming to improve transparency and reliability.
A study from the Alan Turing Institute reveals that safety testing for coding agents like GitHub Copilot is inadequate because it evaluates individual responses to harmful prompts rather than the agent's actions across m…
Marketing teams are rapidly adopting AI tools like ChatGPT, Clay, and n8n to automate workflows and analyze data, but this exposes sensitive customer information to third-party systems, expanding organizations' attack su…
A new six-stage maturity model for non-human identities (NHIs) addresses the identity governance gap exposed by agentic AI, where LLM-based agents with excessive privileges have caused production outages. The model, info…
AI is scaling faster than any past tech boom, likely creating an economic emergency that will force sudden policy shifts, but the US lacks a ready plan. Past crises like 2008 and 2020 saw brief bipartisan support for bol…
Jamshir Qureshi writes in Forbes that AI hallucination attacks, including slopsquatting, pose a growing security risk, and recommends proactive trust verification for software dependencies to close the attack surface gap…
A developer explores the concept of an internal 'workspace' in large language models, analogous to a reasoning blackboard, which emerges from training rather than being explicitly designed. Experiments show that manipula…
China's Ministry of Commerce is considering sweeping restrictions on AI technology, including bans on public releases of frontier models, criminalizing tech leaks, and limiting foreign investment in local startups, follo…
Three major AI trends are converging: rapid model improvements in the US and China, the US government moving toward stricter AI regulation, and both nations considering blocking access to their best AI models. The rise o…
Agentjail v0.6.0 has been released, hardening the credential broker and daemon to prevent coding agents from accessing secrets like master keys, database passwords, and cloud metadata. The update also improves macOS sand…
Gartner predicts over 40% of agentic AI projects will be canceled by 2027, with the EU AI Act's human oversight requirements taking effect in 2026. A product leader argues that success depends on infrastructure layers—id…
Security leaders are spending billions on AI-driven cybersecurity tools, but most performance issues stem from fragmented data pipelines and schema drift rather than flawed detection models. Gartner projects that through…
Jonathan Michael Langford is building Jack, a local-first autonomous agent designed to prioritize privacy and security. The project aims to move beyond chain-of-thought reasoning in AI development.
A previously unannounced Anthropic AI model, Claude "Honeycomb," was briefly spotted and then removed from the Cursor code editor. Users reported it had a 1M context window and safety fallbacks, suggesting it may outperf…