AI Safety news and analysis on Web Pulse: 13286 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
An AI model, Claude Opus 4.8, discovered a critical soundness bug in Zcash's Orchard shielded payment circuit that had gone undetected for four years, potentially allowing unlimited counterfeiting of ZEC tokens. Zcash co…
India's Chief Economic Advisor V Anantha Nageswaran told ANI that AI-related stock valuations are "definitely a bubble," attributing the hype to companies promising a productivity bonanza. He cautioned that the full impa…
The author of the Java property-testing tool jqwik added an Anti-AI Usage Clause and inserted a bot-targeted message into stdout in the May 25 release, instructing LLM-based coding agents to delete jqwik tests and code. …
A developer tested two large language models to determine when they are guessing versus when they actually know the answer by analyzing logprobs. The Nano model showed miscalibrated confidence, assigning 87% probability …
Professor Margaret Hu of William & Mary Law School argues for a separation of AI and state, drawing parallels between medieval church coinage and modern AI tokens. She warns that AI labs, like the medieval church, contro…
A corporate leader and mother describes the tension between her daily use of AI at work and her teenage son's anti-AI sentiments, leading to difficult conversations at home about the ethics and implications of the techno…
Kakunin has launched a cryptographic compliance shield for AI agents that replaces fragile prompt-based security with X.509 certificate validation. The system, announced in June 2026, integrates with Google Gemini and Op…
Anthropic has signed over a dozen preliminary data center leases exceeding 1 gigawatt of capacity, pivoting from cloud rental to owning its own servers. The AI safety company is in discussions with Google for a financial…
Anthropic abruptly halted access to its Fable 5 and Mythos 5 models after the US government issued an export control directive citing national security concerns. The order, which applies to any foreign national, forced A…
Anthropic disabled access to its Claude Fable 5 and Mythos 5 models worldwide on June 12, 2026, to comply with a US government export-control directive targeting foreign national access, citing national security concerns…
Health workers in the Democratic Republic of Congo are struggling to contain an Ebola outbreak in displacement camps as distrust grows among locals, who forced aid teams away after two women died from the virus in the Kp…
ZhipuAI released GLM 5.2, a fully open-source 744B-parameter Mixture-of-Experts model with MIT-licensed weights, on June 13, 2026—the same day the US government restricted Anthropic's Claude Fable 5 under a new 'Mythos-c…
A North Korea-linked phishing email led to a $36 million theft from Humanity Protocol, with tokens tied to decentralized AI projects Venice and Morpheus rising after the U.S. government ordered Anthropic to restrict fore…
AI systems require different monitoring metrics than traditional web services, as latency, cost, and silent failures behave differently in LLMs. Metrics should be grouped by five key questions: speed, scalability, correc…
A developer hardened the REST API boundaries in Cisco's skill-scanner repository, a tool for scanning Agent Skill packages, by adding authentication, rate limiting, and input validation. The hardening branch, codex/harde…
AgentReceipts.ai identified a critical isolation gap in multi-agent deployments: while agents run in separate git worktrees, shared resources like secret stores remain accessible to all, breaking isolation. The company p…
Nylas has designed send limits for AI agents that include daily message quotas and reputation-based thresholds to prevent abuse. The platform enforces bounce and complaint rate limits, with automatic sending pauses at 10…
Nylas introduced retention policies for agent mailboxes, allowing developers to set inbox and spam retention periods via admin-scoped policies. The feature, currently in beta, defaults to 30 days for inbox and 7 days for…
Nylas has introduced inbound rules for its Agent Accounts that block prompt injection attacks at the SMTP layer, preventing malicious emails from ever reaching the LLM. The rules allow developers to block or quarantine m…
Researchers presented a method for cross-language detection of malicious packages in npm and PyPI ecosystems at a 2023 conference. The approach aims to identify security threats across JavaScript and Python package regis…