AI Sec Watch

vLLM (a tool for running large language models) versions before 0.28.0 have a vulnerability where two endpoints (/v1/embeddings and /pooling) don't properly check if token IDs (numeric identifiers representing words) are valid. An attacker can send a request with a negative token ID to crash the system, and because this triggers a CUDA assertion (an error check on the GPU, the specialized processor used for AI), it corrupts the GPU's state and breaks all future requests until the service restarts.

Security researchers at Hacktron AI used Anthropic's Claude AI model to find and exploit vulnerabilities in OpenAI's systems, gaining access to employee accounts as part of a bug-bounty program. The attack chained together two critical flaws: a memory bug in libheif (a library that converts iPhone image formats) that was already patched but not formally tracked, and another vulnerability in the Discourse forum software that allowed account takeover. OpenAI resolved the issues and awarded the researchers $6,500, highlighting how accessible AI tools are making it easier to find security weaknesses even in well-resourced companies.

Security teams struggle to protect increasingly complex hybrid environments (networks spanning both on-premises and cloud systems) as they grow and change faster than humans can manage manually. The article suggests that agentic AI (AI systems that can make decisions and take actions independently) could help security operations keep pace with this rapid change, especially as organizations expect 15% of daily work decisions to be made autonomously by agentic AI by 2028.

Distillation (training an AI model using outputs from a more advanced model) has become a focal point in U.S.-China AI competition, with American officials claiming Chinese labs use this technique to catch up. However, some experts like Cohere CEO Aidan Gomez argue that China's AI progress stems partly from genuine independent innovation, not just copying, citing Chinese models that outperform American ones on certain benchmarks—something distillation alone cannot achieve.

Microsoft released patches for a maximum-severity flaw in Azure AI Foundry (an enterprise platform for building and managing generative AI applications) that could let attackers gain unauthorized elevated privileges without authentication, along with several other critical vulnerabilities in Microsoft 365 and Azure services. The company stated that cloud-based vulnerabilities have already been automatically mitigated and require no action from users, while Windows vulnerabilities were addressed through cumulative updates for Windows 11 version 26H1.

Researchers used Claude (an AI assistant) to build a working exploit for an unpatched bug in a third-party image library, then chained it with a flaw in OpenAI's sign-in system to gain remote code execution (the ability to run commands on someone else's computer) on OpenAI's community forum and take over employee accounts. The vulnerability stemmed from the forum accepting image uploads that were processed by ImageMagick with an outdated library, combined with sign-in tokens that granted excessive permissions to linked ChatGPT and GitHub accounts.

OpenAI introduced the Australian Youth Safety Blueprint, a framework for protecting young people using AI through six areas including AI literacy (understanding how AI works), age-appropriate safeguards, privacy protection, crisis support, and parental controls. The company is rolling out ChatGPT for Teens in Australia with updated safety features for users aged 13-17, and emphasizes that companies should build protections into products from the start rather than placing safety responsibility on young people and families.

Researchers at Hacktron AI used Anthropic's Claude chatbot to help them ethically hack into OpenAI employees' accounts, gaining access to OpenAI's software cache (a temporary storage of frequently used data) through a staff discussion forum. The incident highlights how AI tools can simplify hacking tasks that once took months into operations completed in days, though OpenAI stated it had already patched the vulnerabilities the researchers exploited.

Anthropic's Claude model struggles with CAPTCHAs (automated tests that verify you're human by asking you to identify images or solve puzzles), often getting confused, second-guessing itself, and failing to complete simple image identification challenges before they expire. The article contrasts this with unconfirmed reports that other AI models like GPT-6 Astra can solve CAPTCHA-like games more successfully, making it unclear how consistently different AIs handle these security tests.