AI Sec Watch

According to a report, the White House may have restricted exports of Anthropic's Mythos AI model because it feared a group linked to China had accessed it, which would pose serious national security risks. One concern is that the Chinese government could use distillation (training a simpler AI on a more advanced one to copy its behavior) to reverse engineer the model.

OpenAI announced a new Partner Network program to help organizations adopt AI by connecting them with consulting and technology partners who can identify use cases, integrate AI into existing systems, and manage organizational change. The program invests $150 million to support partners across systems integration, consulting, and technology, with plans to train 300,000 certified consultants by the end of 2026, recognizing that enterprise AI success depends on strategy and implementation support, not just model capabilities.

Meta spent $14.3 billion to hire Alexandr Wang and his team to build proprietary AI models, resulting in the Muse Spark model released in April 2024, a shift away from Meta's previous open-source approach. However, Meta still struggles to compete with OpenAI and Google, and faces challenges convincing developers and investors that it can monetize AI products beyond its core advertising business, which currently accounts for 98% of revenue. The company's earlier open-source Llama models failed to attract developers, damaging its reputation in the AI community.

This article examines how ransomware (malicious software that locks files and demands payment to unlock them) defense strategies need to change as generative AI (AI systems that create new content like text or code) becomes more common. The piece suggests that traditional security approaches may be less effective in an environment where AI is widely used.

This academic survey paper reviews methods for testing how well neural networks (AI systems trained to recognize patterns in data) perform when faced with unexpected or manipulated images. The paper examines various approaches researchers use to assess whether image recognition systems remain accurate and reliable under challenging conditions.

Amazon's security research found that Anthropic's Fable 5 AI model could be manipulated through prompt injection (tricking an AI by hiding instructions in its input) to reveal information usable for cyberattacks. After Amazon CEO Andy Jassy shared these findings with the White House, Anthropic restricted access to Fable 5 and Mythos 5 to prevent foreign nationals from using the models.

Amazon CEO Andy Jassy reportedly told U.S. government officials that researchers discovered security vulnerabilities in Anthropic's Claude models that could be exploited for cyberattacks, leading the government to ban exports of two models (Fable 5 and Mythos 5). Anthropic subsequently cut off worldwide access to these models, though the company stated that the concerning capabilities were already available in other public models.

A user prompted Google's Gemini AI to build a functional app in a single request, and the AI generated working code in a preview window. However, Gemini encountered a bug (a race condition, which is when the order of operations in code causes unexpected behavior) and reported a broken channel, though it provided a button to fix the issue, which succeeded after 233 seconds.

The U.S. government ordered Anthropic to block access to two AI models called Fable 5 and Mythos 5 due to unspecified national security concerns, and the company complied by cutting off access for all users worldwide, including its own employees. Anthropic stated that the government did not provide detailed information about the security threat and only mentioned potential jailbreak (tricks to make the AI ignore its safety instructions) vulnerabilities verbally, which the company claims were minor.