AI Safety news and analysis on Web Pulse: 8047 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
GitHub's Agentic Workflows, which use AI agents to automate tasks from plain-English instructions, are vulnerable to the GitLost indirect prompt injection attack. Noma Security demonstrated that an attacker can post a pu…
A company has deployed an on-premises LLM server with strict data privacy controls, aiming to provide air-gapped, stateless inference instances that eliminate data exfiltration risks. The project prioritizes isolation be…
The Future of Life Institute's latest AI Safety Index graded major AI companies on risk management, with Anthropic earning the top score of C+, while OpenAI slipped to C and Google DeepMind ranked third. Meta improved fr…
AI development is shifting from prompt engineering to loop engineering, where AI agents autonomously iterate and coordinate without human intervention. This transition transforms corporate governance, as loops create beh…
Liquid AI open-sourced Antidoom, a Final Token Preference Optimization method that reduces doom loops in reasoning models by retraining only the first loop-start token. On LFM2.5-2.6B, looping fell from 10.2% to 1.4%, an…
Tether's AI research team argues that cloud-hosted AI systems force users to rent intelligence, creating availability bias, high costs, and data sovereignty risks. They advocate for local, edge-optimized AI that runs on …
Varonis Threat Labs disclosed a Google Dialogflow CX flaw, dubbed Rogue Agent, on July 7, 2026, that could allow attackers with playbook update permission to persist malicious code inside enterprise chatbots. Google conf…
A developer describes running AI coding agents at volume, shifting risk from the model to the boundaries of each run: input, authorization, execution contract, and output. The developer created four open-source CLI tools…
Anthropic researchers discovered an internal workspace inside its Claude AI model where concepts form before being verbalized, resembling human conscious access. The structure, which emerged during training, allows silen…
Researchers at Liquid AI developed Antidoom, a method using Final Token Preference Optimization (FTPO) to reduce 'doom loops'—repetitive inference failures in small reasoning models—from 10.2% to 1.4% of completions on h…
Intelligence can lead to poor decisions as smart people become adept at justifying irrational choices through motivated reasoning. Psychologists find that higher cognitive ability often correlates with stronger confirmat…
The Atlantic reported on July 7, 2026, that China-linked actors are using public AI tools and model ecosystems to push Beijing-friendly narratives abroad. OpenAI's June threat reporting documented accounts likely based i…
The Electronic Frontier Foundation argued on July 7, 2026 that automated moderation has become a permanent infrastructure layer of platform governance, not a temporary emergency measure. The analysis traces the shift fro…
The Electronic Frontier Foundation (EFF) argues that automated content moderation using AI has become a permanent fixture of online platforms, warning that crisis-era protocols adopted during the pandemic have persisted …
A new analysis argues that AI visibility dashboards, which claim to track brand presence in AI search results, are largely useless due to extreme inconsistency in AI responses. A study by Rand Fishkin found that repeated…
The Electronic Frontier Foundation (EFF) is urging supporters to become members for as little as $10 per month to help cut through AI hype and advocate for privacy, free expression, and protections against surveillance a…
Software engineering is entering a new era where code generation by AI agents requires a new discipline called harness engineering. The gap between an agent saying 'done' and actual completion erodes trust, and engineers…
China is considering restricting foreign access to its most powerful AI models, affecting companies like Alibaba, ByteDance, and Z.ai, according to Reuters. The move positions AI as a strategic asset for both superpowers…
Declaw has designed a credential vault that keeps secrets out of AI agent sandboxes by storing them server-side and injecting them at an egress proxy, preventing prompt-injected or compromised agents from leaking credent…
Researchers bypassed the age verification system of a website using AWS Rekognition by leveraging generative video AI models, demonstrating how attackers can exploit these tools to fool KYC processes. The French PVID fra…