AI Safety

A developer spent weeks auditing smart contracts for a Web3 banking project, SovereignBank Web3, using five top AI models individually—Claude Opus, Gemini Ultra, ChatGPT, DeepSeek, and Grok—through thirteen rounds, with …

A developer nearly caused a destructive rollback of a WordPress plugin repository after adding an overly permissive `Bash(git *)` rule to Claude Code's settings, which silently allowed commands like `git reset --hard HEA…

Amit Gautam, CTO at AI-security vendor Abluva, outlined in a Help Net Security video how autonomous AI agents introduce governance and security risks inside enterprises. Gautam demonstrated a scenario where a reconciliat…

AgentSafeLabs tested Claude Haiku against prompt injection attacks, with two of three ASI01 tests passing and one returning UNCERTAIN. The UNCERTAIN result highlights challenges in defending agentic systems against indir…

Companies that provide developers with laptops using their full names and usernames, then deploy coding agents like Gemini, expose developers' personally identifiable information (PII) to U.S. data centers without protec…

An engineer who analyzed the EU AI Act's Article 17 found that most AI agent teams are unprepared for the August 2, 2026 compliance deadline, particularly regarding quality management system (QMS) requirements. The engin…

An Anthropic employee expressed confusion about their role, stating that on good days they feel automated work renders their efforts meaningless, while on bad days they cannot understand why systems break. The employee's…

A developer has built a "Talk Mode" extension for the Pi coding agent that prevents the AI from editing, creating, or modifying files. The extension, toggled with Shift+Tab, injects a system prompt that restricts the age…

Japanese Digital Minister Hisashi Matsumoto warned Friday that Japan risks becoming an "AI colony" if it fails to keep pace with rapid artificial intelligence development. Matsumoto made the statement while defending a b…

South Korea will require all internet communities and forum owners to scan every user-uploaded image and video using AI censorship tools, starting July 1st, under revised telecommunications laws. Website operators must p…

A developer working on the CLAIM-24 self-correcting systems project has demonstrated a critical failure mode in AI agent authorization: a timestamp-only gate allowed an agent to send data to a partner whose access had be…

Amit Gautam, CTO at Abluva, warned that autonomous AI agents pose escalating security risks as their numbers surpass human oversight in enterprises. He cited a real incident where a financial firm's reconciliation agent,…

James Deaker, head of consultancy Korukea Media, warned that autonomous AI systems making commercial decisions without clear governance could pose significant risks to the digital advertising industry. Citing examples li…

A new report from Ivanti reveals that 68% of IT professionals have personally observed AI hallucinations in their operations, even as autonomous AI systems increasingly take direct action in enterprise environments. The …

Meta Platforms Inc. Chief AI Officer Alexandr Wang said health capabilities will be a key differentiator for future Meta AI models as the company scales them to billions of users. Wang, who joined Meta from Scale AI afte…

Enterprises are pushing autonomous AI agents into production for coding, transaction processing, and customer service faster than their identity and access controls can keep up, according to a report from Economic Times …

A new study from researchers at Northwestern, Stanford, Cornell, and All Hands AI reveals that frontier LLM agents waste 28–64% of tokens on tasks they are doomed to fail. The paper, BAGEN: Are LLM Agents Budget-Aware?, …

Researchers have introduced MCBench, a new benchmark designed to assess the safety of Omni Large Language Models that process vision, audio, and text. The benchmark includes 1,196 scenarios across four safety categories,…

Researchers at a major AI lab have identified a specific neural subgraph within a large language model (Qwen3-4B-Instruct-2507) that governs how the model trades off short-term gains against long-term consequences. The t…

Researchers have developed a cross-model safety steering framework that transfers a safety direction from a source large language model to a target image or video generator using only benign data, eliminating the need fo…