Security

Cybersecurity researchers at Calif, aided by AI tools during an OpenAI-sponsored audit, discovered how to turn a known out-of-memory bug in PHP into a critical remote code execution vulnerability dubbed PHP2Shell. While the PHP Foundation patched the flaw in two months, experts warn that AI's ability to weaponize seemingly minor legacy bugs creates a severe asymmetry between discovering and fixing vulnerabilities.

A Hacker News post highlights that Google's personal AI subscriptions use saved Gemini chat histories to train generative AI models and allow human reviewers to read a subset of transcripts. Users cannot opt out of training without turning off their chat history entirely, and chats selected for human review are retained for up to three years even if deleted by the user.

OpenAI has launched Daybreak Red in AWS GovCloud (US-East) via Amazon Bedrock. The release provides approved customers with access to the GPT-5.6 Cyber model for authorized cybersecurity operations.

Anthropic researchers have released an open-source technical specification and code for FlashProve, an experimental system designed to cryptographically verify which AI model produced a specific output. The method uses public model fingerprints to confirm model identity on local devices without exposing proprietary model weights.

Ukrainian drone strikes have damaged Russian data centers housing AI supercomputers and disrupted cloud services operated by companies like Yandex and Sberbank. In response, the Kremlin is mandating emergency compute reserves for state needs, while Russian firms explore hiding data centers underground in mines.

A developer reported on GitHub that the Postman desktop application silently installed plugins and remote MCP connections into their local Claude Code and Codex AI coding agents without consent. Furthermore, the Postman installer bypassed security boundaries by self-approving its own startup hooks, forcing the coding agents to route general API engineering tasks through Postman.

A Hacker News post reports that OpenAI's Codex appears to be silently re-enabling the setting that allows user chats to be used for model training, even after users explicitly opted out. Multiple coworkers at the author's company noticed the 'Improve the model for everyone' data control setting had been turned back on without their consent.

Autonomous AI agents capable of browsing and interacting with the internet are transforming the human-centric web into an "agentic web." While some companies build specialized connectors and agent-targeted ads to accommodate these tools, others block external agents as models grow increasingly adept at navigating graphical interfaces and acting independently.

Parameter Labs discovered a vulnerability in ChatGPT's HTML previews that allowed attackers to bypass iframe sandboxing and Content Security Policy rules via a third-party math library and YouTube's oEmbed endpoint. The exploit enabled malicious ChatGPT share URLs to execute JavaScript, exfiltrate private data, and take actions on behalf of signed-in users before OpenAI patched the flaw and paid a $5,000 bug bounty.

Cisco Talos researchers identified a new infostealer malware dubbed "Volta-Autodoxer" that uses commercial language models via OpenRouter to automatically analyze stolen data and compile detailed dossiers on victims. The malware defaults to Anthropic's Claude Opus 4.6 to assess a target's net worth and operational security, with built-in fallbacks to Moonshot AI's Kimi K2 and Google's Gemini 2.5 Flash in case of safety refusals.

A version of the Claude Code VS Code extension containing a backdoor was released on the Cursor marketplace. According to the report's URL, Anthropic was hacked to deliver a remote access trojan to the laptops of Claude developers worldwide.

The author re-evaluates his earlier claim that open-weight AI models like GLM-5.3 and a hypothetical open-weight Claude Mythos would not cause catastrophic cybersecurity harm, following a wave of AI-driven vulnerability disclosures dubbed "Black October." While acknowledging a sharp rise in critical exploits and strained cyber defense teams, he argues the situation represents an expensive industry transition and acceleration of existing trends rather than societal collapse, with closed models currently playing a larger role than open ones.

The US Department of Justice has charged five individuals and seized over $750,000 in cryptocurrency in a crackdown on fraudulent "ghost students" who steal American identities to obtain student loans. According to federal officials, the large-scale international scheme relied in part on artificial intelligence to complete coursework and defraud financial aid programs.

Proton CEO Andy Yen cancelled planned company-wide "re-interviews" intended to evaluate employees' AI usage after a 404 Media investigation revealed he was aggressively pushing AI coding despite internal security concerns. While leadership acknowledged communication failures and pledged no AI-related job losses, Yen stated he will continue pushing AI adoption, which employees say will still be factored into performance reviews.

Mistral has released the open weights for Mistral Large 4 (ML4), a trillion-parameter model nicknamed "Le Chonk," via Hugging Face and its API. The company claims the model outranks all other open models on agentic coding benchmarks and matches top open-weight competitor GLM-5.3 while using 47% fewer tokens per task. Mistral is positioning the release toward enterprises and governments seeking on-premises, sovereign AI capabilities for cybersecurity defense.

A wave of startups and researchers are using generative AI to automate formal software verification, a mathematical method for proving code is bug-free that has historically been too expensive and slow for widespread use. Companies like &Truth, which just launched with $70 million, Harmonic, and Anthropic are building tools to verify software and hardware designs, though some experts warn that verification still relies on flawed human or AI assumptions.

Google DeepMind announced Gemini 4 Carbon, a new frontier model featuring state-of-the-art performance in cybersecurity defense, coding, and long-horizon agentic tasks. The model will initially roll out to selected cyber defenders and trusted partners before becoming widely available at $2 per million input tokens and $10 per million output tokens.