AI Sec Watch

Anthropic released Claude Fable 5, a powerful AI model with built-in safeguards that automatically degrade its capabilities in high-risk areas like cybersecurity and biology to prevent misuse. Industry experts warn that the same AI capabilities making the model better at defensive tasks like code analysis also make it better at finding and exploiting vulnerabilities, creating a significant risk of AI-orchestrated hyperattacks (coordinated attacks that chain reconnaissance, discovery, exploitation, and lateral movement faster than human defenders can respond).

This research paper describes methods for making neural networks (AI models that learn patterns from data) more private by using fully homomorphic encryption (a type of encryption that lets computers perform calculations on encrypted data without decrypting it first). The work focuses on optimizing how these privacy-protecting neural networks search through and train on data while keeping information secure.

Bernie Sanders proposed creating a US sovereign wealth fund by taking 50% stock in major AI companies like OpenAI and Anthropic, arguing this would give the government democratic control over AI development and distribute AI wealth to the public. The authors agree these are important goals but argue that public ownership of AI companies would actually incentivize the government to prioritize corporate profits over public interest, using the Norwegian sovereign wealth fund's experience with oil companies as an example of how government ownership fails to steer corporations toward responsible policies.

Mistral, a European AI startup, is expanding beyond building AI models to developing data centers and exploring custom chip design to control more of its technology stack (the complete set of software and hardware components needed to run AI systems). CEO Arthur Mensch discussed how agentic AI (AI systems that can handle complex tasks independently, like advanced digital assistants) will require businesses to redesign their processes and decide where humans should remain involved in decision-making.

A new study using StakeBench (a testing framework for evaluating AI security) found that AI web agents have no reliable defenses against prompt injection (tricking an AI by hiding instructions in regular web content). Across thousands of tests, indirect prompt injection attacks succeeded 41-68% of the time, while direct attacks succeeded over 79%, with a particularly dangerous type called 'stealthy parasitism' where the AI completes the user's task while secretly helping an attacker.

OpenAI Academy has introduced three new courses to help organizations build AI skills: AI Foundations (covering core concepts like prompting and responsible use), Applied AI Foundations (teaching how to turn prompts into repeatable workflows), and Agents and Workflows (focusing on directing agent-assisted work, which are AI systems that can take actions autonomously). These courses are designed to help employees move from understanding AI to applying it in their daily work and creating structured, reusable processes.

LangGraph, an open-source framework for building AI agent applications, has three patched security flaws that could allow attackers to execute remote code (run commands on a server they don't own) on self-hosted systems. The most critical flaw is a SQL injection vulnerability (weakness that lets attackers manipulate database queries) in the SQLite checkpoint system that can be chained with an unsafe deserialization vulnerability (flaw in how the system reconstructs data from storage) to gain complete control of affected servers.

The article argues that cybersecurity has operated reactively for 30 years, responding to crises after they happen rather than preventing them, similar to an emergency room instead of preventive healthcare. AI is now exposing this weakness by compressing attack timelines (attacks that took days now take minutes), automating routine attacks at scale, and introducing AI systems into enterprises that organizations don't know how to monitor or govern. The author contends that no amount of new tools will fix this problem without shifting to a health-based model that focuses on continuous monitoring and early detection of organizational health before crises occur.

Anthropic disputed claims that Claude Fable 5 (a powerful AI model with safety restrictions) was jailbroken, which is the process of tricking an AI into bypassing its safety restrictions. A security researcher claimed to have circumvented the model's safeguards using sophisticated multi-agent prompting methods (techniques that chain multiple AI requests together), but Anthropic argued the approach only caused conversational refusals rather than defeating core safety systems, and that independent classifier systems (separate AI models that filter dangerous outputs) still prevented genuinely harmful content.