AI Sec Watch

Trail of Bits has published a new chapter in their Testing Handbook that teaches security testing techniques for Rust programs, covering both dynamic analysis (testing while code runs) and static analysis (examining code without running it). The guide includes information about Rust's security guarantees, specific tools like Clippy (a code linter) and Miri (which detects undefined behavior, or code that doesn't follow language rules), common security mistakes to watch for, and a new Claude Code plugin called rust-review that automatically checks Rust code for over a dozen types of security bugs.

A flaw in the vllm-orchestrator-gateway component (a system that manages requests to AI language models) logs sensitive information like authorization headers (credentials that prove who you are), bearer tokens (temporary access passwords), and full chat conversations to persistent logs (files stored on disk). Any user with access to the logging system can read this data, which may include personally identifiable information (PII, such as names or account details) and secrets, allowing attackers to steal credentials and private conversation content.

Google is adding AI features to the Waze navigation app, including integration with Gemini (Google's AI assistant) to help drivers personalize their trips. Two of the four new updates use Gemini: an updated conversation reporting feature that lets drivers use voice commands to report traffic incidents and map updates, and a new Destination Search feature that also uses voice commands to help find nearby places like coffee shops.

Organizations often create AI risk registers (documents listing potential AI problems and their severity) but lack actual incident response plans for when AI failures occur in real workflows. The source explains that risk registers provide visibility into problems but cannot preserve evidence, notify leaders, or decide whether to shut down a system—and AI incidents are harder to recognize than traditional security breaches since they may appear as bad recommendations or data exposure rather than obvious attacks.

AI is dramatically speeding up cybersecurity work at large organizations like AWS, where vulnerability detection and fixes that once took months now take minutes to hours. However, security experts warn that AI could worsen an existing divide between wealthy organizations with resources and expertise versus smaller organizations (like rural hospitals, community banks, and local governments) that lack the money, staff, and time to adopt AI tools, potentially deepening a cybersecurity inequality that has existed for over a decade.

A vulnerability (CVE-2026-15531) exists in HashNeRF-pytorch's checkpoint file handler that allows unsafe deserialization (converting data back into code objects) when loading files through the torch.load function. An attacker with local access can manipulate the checkpoint file path to execute arbitrary code, and the exploit is publicly known.

OpenAI temporarily removed the five-hour usage limit for GPT-5.6 Sol (a powerful AI model for coding and complex tasks) on its Plus, Pro, and Business plans after demand surged. The company also made GPT-5.6 Sol more efficient so it uses fewer tokens (the basic units the model processes) per request, and reset everyone's usage counter to give users more capacity.

Anthropic has extended free access to Claude Fable 5 (a powerful AI model) for paid subscribers until July 19, 2026, allowing users to access it at no extra cost up to 50% of their weekly usage limits. Previously, the deadline was July 7, then extended to July 12, and now extended again to July 19. After hitting the 50% weekly allowance, users can either pay for additional usage credits or switch to other Claude models.

This article covers a public dispute between Elon Musk and Sam Altman on X (a social platform Musk owns) following Apple's lawsuit against OpenAI over alleged trade secret theft. The conflict stems from a long-standing disagreement over OpenAI's transformation from a nonprofit to a for-profit structure, which Musk lost in court earlier in 2026, and intensified as both leaders promoted competing AI models (Musk's Grok 4.5 and Altman's GPT-5.6 Sol).