
Security firm Apiiro found that Claude Code exposes secrets in public repositories because it stores 'Don't Ask Again' permission approvals in local files on developers' machines. Apiiro identified 1,122 secrets in claude/settings.local.json files, plus another 1,160 in git history, the shared settings.json, and agent chat transcripts committed to repos.
Amazon reports that China-linked threat actors used AI agents to exploit a zero-day vulnerability in medical software. The report shows attackers automating parts of an intrusion with AI agents, targeting healthcare-related software.
Google announced Agent Compute Engine, a managed service that runs AI agents in isolated microVMs. It adds checkpointing so execution environments can be suspended and resumed. Credentials, API keys, and OAuth refresh tokens stay outside the agent sandbox, and an external Agent Gateway authenticates requests and injects the secrets they need.
AWS Lambda now runs Event Functions on EC2 G-series GPU instances through Lambda Managed Instances, with up to 48 concurrent invocations sharing one model per GPU. Prebuilt vLLM and SGLang base images need no handler code, and asynchronous invocations can run for up to 90 minutes. Billing is the instance price plus a 9.75 percent fee, and early hands-on tests report an eight-minute publish time and no response streaming.
Major US financial institutions are prohibiting third-party bots from logging into customer accounts. They also state that customers bear sole responsibility for costly mistakes made by those bots.
AI spending is moving from large training clusters to smaller, quickly built sites, and modular data centers are emerging as a natural home for inference. Power, cooling, and memory supply now shape where tokens get generated as much as raw chip performance does. Inference-first chipmakers such as Positron AI are positioning to capture that shift.
A New York Times report says Chinese intelligence services were focused on advanced AI chips well before the recent Nvidia-linked scandal. The source material is a brief headline and teaser, so the specific agencies, methods, and evidence are not confirmed here. The story signals that AI hardware has been a sustained target of state-linked espionage.
South Korea's education minister, Choi Kyo-jin, resigned after an AI-assisted cheating scandal surrounding the country's national university entrance exam, the suneung. The scandal began when the top scorer in a national tax accountant exam admitted using a miniature camera and smartphone to achieve a near-perfect score just four days before the suneung. Choi said he 'failed to prepare adequately' despite this year's measures, which included banning phones and checking for cameras hidden in clothing.
A New Stack review compares OpenAI's GPT-6.2 Sol with Anthropic's Claude Opus 5.5 based on the author's own tests. The author finds Claude Opus 5.5 has a mixed record, while GPT-6.2 Sol is less accurate overall but may still offer better value. The excerpt is brief, so the full article would be needed for specific benchmark results and pricing.
The New Stack ran a head-to-head comparison of Anthropic's Fable 5.2 and OpenAI's GPT-6.2 Astra, reporting that Fable 5.2 was slightly more accurate in the author's tests. The piece notes that Astra's launch drew mixed reviews, but says it performed well in the author's evaluation. The excerpt is truncated, so the full benchmark methodology and scores are not available here.
A ScienceDaily report says researchers have found that AI systems may not operate the way the human brain does, challenging a common comparison in the field. The source text offers only the headline and a brief excerpt, so the specific methods and findings could not be verified here.