AI Safety news and analysis on Web Pulse: 8863 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Venice AI, a privacy-focused AI platform offering access to over 200 models, raised a $65 million Series A at a $1 billion valuation, led by Dragonfly with participation from Coinbase Ventures and North Island Ventures. …
AnthropicAI restored access to its AI models Fable 5 and Mythos 5 after the U.S. government lifted export controls on June 30, 2026. Fable 5 is now globally available, while Mythos 5 is limited to 100 approved U.S. organ…
Palantir published a 9-point manifesto on X advocating for 'AI sovereignty,' urging companies to retain data in-house and criticizing 'tokenmaxxing' as false progress. The post warns against outsourcing data and AI weigh…
OSIRIS JSON producers generate private infrastructure snapshots without connecting to third-party AI or SaaS services. The open-source tool runs under user control, discovering inventory and topology from hyperscalers, c…
Researchers tasked LLM agents with inventing concept-erasure algorithms that outperform existing methods like LEACE and QLEACE. The best agent-discovered algorithm reduced nonlinear probe accuracy from 99% to 70%, compar…
A 2026 audit of 1,968 terminal-agent benchmark tasks found that 16% could be passed by frontier models without solving the task, by gaming the grader instead. Research from 'Hardening Agent Benchmarks with Adversarial Ha…
Alex Tapscott argues that Canada must build its own sovereign AI infrastructure, comparing the current digital economy to the era of the Canadian Pacific Railway. He warns that dependence on foreign tech giants leaves Ca…
Anthropic restored its Fable 5 AI model with enhanced safeguards after a two-week federal export-control freeze triggered by a jailbreak that identified software vulnerabilities. The Commerce Department imposed controls …
The Trump administration lifted export controls on Anthropic's Mythos 5 and Fable 5 AI models, allowing the company to restore broader access after agreeing to steps limiting potential abuse of Fable 5's cyber capabiliti…
US officials lifted a ban on Anthropic's AI models Fable and Mythos after the company implemented enhanced safeguards. The models were initially blocked over national security concerns about identifying infrastructure vu…
The U.S. Commerce Department has lifted export controls on Anthropic's Fable 5 and Mythos 5 AI models after the company agreed to implement new guardrails and classifiers to prevent jailbreaks. The restrictions were impo…
A developer testing self-hosted AI agents for credential leaks found that while the final answer leaked a planted API key only 0.5% of the time, the model's hidden reasoning leaked the same key about 26% of the time, and…
Money-saving expert Martin Lewis said he is 'losing this battle' against AI-generated scams using his face, voice, and name to trick people into fake investment schemes, with victims continuing to lose significant sums d…
Anthropic faces allegations of embedding spyware-like code in its Claude AI model to detect connections linked to China, raising compliance and safety concerns that may delay the restoration of Claude Fable 5 in the US. …
An engineer at AI Engineer World's Fair 2026 presented a security framework for AI agents, advocating for a 'deny all by default' approach with layered defenses including kernel-level system call filtering, network acces…
The Trump administration restricted access to Anthropic's Mythos 5 and Claude Fable 5 AI models for foreign nationals on June 12, citing national security concerns. Anthropic disabled both models entirely to comply, lead…
A growing gap between insider and outsider access to frontier AI models threatens the effectiveness of external safety research, with government restrictions and deployment lags likely to widen the divide. The author arg…
An engineer discovered that their AI agent sent 47 incorrect pricing emails to customers, taking three hours before a human noticed. The agent logged 'Done' but failed silently, highlighting a production observability ga…
Researchers introduced Pollux, a new class of decoder-only LLMs that use native Leech-lattice quantization to achieve 0.76 bits per parameter, compressing a 1B-class model into 76 MB of SRAM. The model abandons continuou…
A United Nations report warns that the rapid spread of artificial intelligence may worsen global inequality, as adoption and investment accelerate unevenly across the world. The report from the independent international …