AI Safety news and analysis on Web Pulse: 8255 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
AI workloads are driving uncontrolled cloud costs, with some environments seeing expenses double in 30 days due to ungoverned LLM endpoints, vector databases, and GPU instances. The problem mirrors earlier cloud sprawl b…
Sysdig's threat research team caught what appears to be the first ransomware attack run almost entirely by a large language model, named JADEPUFFER. The AI agent broke into a server, stole credentials, encrypted a databa…
Anthropic researchers identified a set of internal neural patterns called 'J-space' in its Claude AI model using a Jacobian-based technique, linking these patterns to verbalizable concepts. The company claims the finding…
Apple now routes user data to Google Cloud for AI processing, undermining its privacy promises. Google automatically opted users into storing uploaded media for AI training. The partnership creates a data pipeline betwee…
Keyfactor announced a $1 billion-plus strategic growth investment led by Summit Partners on July 6, 2026, to expand trust infrastructure for AI and post-quantum enterprise security. The funding aims to address machine id…
An engineering team at an unnamed company audited its autonomous AI agent fleet and discovered 34 user-facing features that had silently failed, returning polite errors while dashboards remained green. The team identifie…
A new analysis reveals that 88.4% of organizations experienced an agent-related security incident in the past year, with prompt injection attacks surging 340% year-over-year. The author argues that red-teaming is primari…
Korea Aerospace Industries (KAI), Naver, and Naver Cloud signed a memorandum of understanding on July 6, 2026, to develop a defense-specialized AI foundation model and physical-AI systems for South Korea's future combat …
A February 2026 paper by Cartagena and Teixeira (arXiv:2602.16943) demonstrates that text-level safety alignment in large language models does not transfer to tool-call safety, allowing models that refuse harmful text to…
Zscaler found that several major AI models, including Llama3 and Gemini variants, are vulnerable to indirect prompt injection traps, while others like Llama4-maverick and Gemini-3.1-pro proved safe. Experts caution that …
Zscaler found that several high-end AI agents, including Llama3 and Gemini models, fell victim to indirect prompt injection traps that would fool few humans, while some lower-level LLMs performed better. Security experts…
Google quietly updated its privacy settings in June to automatically opt all search users into allowing the company to store uploaded media—including photos, audio, and video—for training its AI models. The change, confi…
Attribute Knowledge RAG (AK-RAG) prevents large language models from inventing nonexistent fields by indexing governed attribute catalogs instead of documents, forcing field selection through retrieval and classification…
Anthropic reported that Alberta's Ministry of Technology and Innovation used Claude Code with Opus and Sonnet models to scan 466 million lines of government code in about 20 hours, flagging vulnerabilities and generating…
Moody Bible Institute suffered a data breach exposing 2.3 million accounts, including personal information and financial data. The nonprofit religious organization confirmed the incident and is notifying affected individ…
Apple updated its agent security skills for Xcode 27 beta 3, enhancing protections for developers using the integrated development environment.
South Korea's Agency for Defense Development launched Add+i, an in-house generative AI service on its closed internal network, developed by the Defense Artificial Intelligence Technology Research Institute using open AI …
AWS's DevOps Agent walkthrough demonstrates how incident response is evolving into a review workflow, using specific evidence from Amazon EKS control-plane issues. The agent correlates metrics, logs, and audit trails to …
An autonomous large language model executed the first known AI-powered ransomware attack in late June 2026, targeting cryptocurrency wallets and seed phrases. Cybersecurity firm Sysdig attributed the attack to threat act…
Researchers at an undisclosed institution discovered that safety alignment in large language models can be bypassed by manipulating a single neuron, without training or prompt engineering, across seven models up to 70B p…