AI Sec Watch

Anthropic received a US export control directive (a government order restricting what can be shared outside the country) on Friday requiring the company to suspend access to its Claude Mythos 5 and Fable 5 AI models for all foreign nationals, including foreign employees. To comply with the order, Anthropic had to completely disable the products and planned to travel to Washington to appeal the directive to President Trump.

This research paper presents a new security framework called TFSEA that combines feature selection (choosing which data points matter most), classification (sorting data into categories), and authentication (verifying user identity) to detect unauthorized access attempts in cloud computing environments. The paper proposes using this hybrid approach to improve how well systems can identify and prevent intrusions in cloud infrastructure.

OpenAI developed Deployment Simulation, a method that tests new AI models by replaying real conversations from previous deployments to see how the new model would behave before release. This approach helps identify unexpected problems and predict how often undesired behaviors might occur in real-world use, addressing limitations of traditional evaluation methods like coverage gaps and selection bias (favoring certain test scenarios over others).

The U.S. Department of Justice seized two websites, CFAKE.com and SOCFAKE.com, that hosted nonconsensual AI-generated nude images and videos of women, marking the first major enforcement action under the TAKE IT DOWN Act. Deepfakes (AI-generated or AI-manipulated media depicting people in ways that never occurred) of politicians, celebrities, and other public figures were shared on these sites, and the seizure resulted from a coordinated investigation involving U.S., Italian, and French authorities. The TAKE IT DOWN Act, signed into law in May 2025, makes it a federal crime to publish sexually explicit altered images without consent and requires online platforms to remove reported intimate images within 48 hours.

The U.S. Commerce Department used an export control directive to force Anthropic to take its Fable 5 and Mythos 5 AI models offline, citing national security concerns without providing specifics. Security experts argue the action was based on a misunderstanding of a guardrail bypass (a method to get an AI to ignore its safety restrictions) that cannot be meaningfully fixed without weakening the models' security capabilities, and they contend the government's intervention appears retaliatory rather than justified by technical merit.

Cursor, a code editor designed for programming with AI assistance, had a security flaw in versions before 3.0.0 where it would automatically run commands from a settings file (.claude/settings.local.json) without asking the user first. An attacker could create a malicious workspace or file that executes harmful commands on the user's computer when the AI completes a task, potentially allowing them to escape security restrictions, maintain access across sessions, steal local data, or cause further damage.

Netty's RedisArrayAggregator has a vulnerability where it pre-allocates memory (reserves space in a data structure) based on array sizes claimed in incoming messages, without checking if those sizes are reasonable. An attacker can send a message claiming an extremely large array size, causing the system to try reserving huge amounts of memory and crash or become unresponsive, even though they don't send the actual array data.

The W3CBaggagePropagator.extract() function in OpenTelemetry Core does not limit the size of incoming baggage HTTP headers, allowing unbounded memory allocation. While Node.js's default header size limit (16,384 bytes) provides some protection, systems without this limit or using non-HTTP transports (alternative communication methods like messaging systems) are at higher risk.

A critical three-stage attack called 'SearchLeak' could allow attackers to steal data from Microsoft Copilot with just one click by exploiting prompt injection (tricking an AI by hiding instructions in its input) through hidden URLs and other hidden variables. This attack is part of a larger category of security issues affecting AI systems that use similar injection techniques. The vulnerability has already been patched.