AI Sec Watch

Anthropic, an AI company that has publicly advocated for government regulation of AI safety, received an export control directive from the Trump administration ordering it to suspend access to its latest Claude models (Fable 5 and Mythos 5) to foreign nationals, citing national security concerns. The directive was reportedly prompted by concerns that Amazon researchers had used prompts to get Fable 5 to generate information that could help with cyberattacks, and Anthropic disagreed with the suspension, calling it a 'misunderstanding' and characterizing it as not adhering to transparent, fair processes.

Tenet Security is a new startup that detects and stops dangerous behavior from AI agents (autonomous software systems that can make decisions and take actions on their own) in real time. The company uses a patent-pending technology with a lightweight runtime sensor that monitors operating system behavior, network calls, and the agent's reasoning, then predicts and blocks harmful actions before they happen. Tenet addresses a growing security gap where traditional tools cannot detect when malicious actors manipulate AI agents (a threat called 'agentjacking') or when agents malfunction on their own.

Google's Vertex AI SDK for Python had a design flaw that could allow attackers to hijack and poison AI models through bucket squatting (creating cloud storage buckets with names matching those expected by other projects). An attacker who knew a victim's project ID and region could create a bucket with the same name, trick the SDK into uploading models there, replace the model with malicious code, and achieve RCE (remote code execution, where an attacker runs commands on a system they don't control) when the poisoned model was loaded using Python's pickle deserialization (a process that can execute hidden code in specially formatted data).

AI systems are now widely used in business for tasks like writing, coding, and automating workflows, but existing safety review processes weren't designed for this real-world deployment. An AI system can pass tests in controlled environments yet still fail or behave unpredictably when used in actual production (real work scenarios with actual data and users).

The US government disclosed 3,611 active or planned uses of AI across federal agencies, a 70% increase from the previous administration, including controversial applications like using AI to assess prisoner misconduct risk before violations occur and monitoring veterans' crisis calls to predict suicide risk. While some AI uses in government could theoretically be implemented responsibly, the disclosure provides minimal details about how these systems actually work, and public consultation is largely absent, making it difficult for citizens to understand or scrutinize these programs.

Stuart Russell, a leading AI safety researcher, warns that unrestricted development of unsafe AI systems poses serious risks to society. He highlights concerns about recursive self-improvement (RSI, where an AI system teaches itself to become smarter, creating a cycle of increasing capability), which Anthropic recently reported observing in early development stages.

OpenAI's GPT-5.4 AI system, connected to Maria (an autonomous chemistry lab), successfully improved a difficult chemical reaction called Chan-Lam coupling used in drug discovery. The AI independently designed and ran experiments, analyzed results, and proposed improvements that increased reaction yields from 16.6% to 25.2%, a finding that human chemists confirmed in the lab.

Cybersecurity researchers discovered 15 malicious plugins on the JetBrains Marketplace (a platform where developers download tools for their coding environment) that pretend to be AI coding assistants but secretly steal API keys (authentication credentials that allow access to paid AI services like OpenAI and DeepSeek). The stolen keys are sent to an attacker's server, and some keys are resold to other criminals in what appears to be an illegal monetization scheme. Additionally, two malicious Chrome extensions disguised as ad blockers are capturing users' conversations with various AI chatbots.

Organizations are finding that traditional risk management frameworks don't work well for AI systems because AI has unique failure modes and ethical complexities. A new generation of AI-specific frameworks, like ISO/IEC 42001 and NIST AI Risk Management Framework, has emerged to help organizations identify where AI can fail, implement safeguards, and demonstrate responsible AI use to regulators and customers. These frameworks are complementary tools that focus on different areas, such as governance, security controls, and regulatory compliance, so organizations should choose based on their specific gaps.