AI Sec Watch

Researchers discovered that large language models (LLMs, AI systems trained on massive amounts of text data) frequently generate fake web domain names that don't actually exist but sound like they belong to real companies. Attackers are registering these made-up domains (a practice called phantom squatting) to intercept traffic sent by AI systems and users who trust the LLM's output, creating a new supply chain attack vector (a way to compromise software development and deployment). The researchers identified over 13,000 confirmed malicious URLs and approximately 250,000 unregistered hallucinated domains that attackers could exploit.

Anthropic has announced that the Department of Commerce lifted export controls on Claude's powerful AI models, Fable 5 and Mythos 5, and will begin restoring access to Fable 5 on Wednesday. The company is also rolling out identity verification (KYC, or know-your-customer checks) for certain Claude features, requiring users to provide government-issued photo ID and a live selfie through a third-party verification partner called Persona to prevent misuse and comply with legal requirements.

Anthropic's Claude Fable 5 AI model is being brought back online after the U.S. Department of Commerce lifted export controls that had previously blocked it. The company plans to restore access to users globally starting Wednesday across Claude platforms and cloud services like AWS, Google Cloud, and Microsoft Foundry, though without a specific timeline for completion.

Microsoft SharePoint Server has a deserialization of untrusted data vulnerability (a flaw where the software unsafely processes data from an untrusted source, allowing an attacker to inject malicious code), which lets an authorized attacker run code over a network. This vulnerability is actively being exploited in real attacks. Organizations must apply security updates following CISA's BOD 26-04 guidance by July 4, 2026, or stop using the product if no fix is available.

Anthropic announced that the U.S. Department of Commerce has removed export restrictions on two AI models, Claude Fable 5 and Mythos 5, allowing the company to restore user access to these models starting the next day. This announcement suggests that government controls on advanced AI technology exports had previously limited who could use these models outside certain regions.

This release document describes updates to security techniques and mitigations related to AI systems, including new attack methods like stealing web session cookies and jailbreaking LLMs (large language models, AI systems trained on massive amounts of text), as well as updated defenses like AI guardrails (safety features that prevent unwanted outputs) and telemetry logging (monitoring system activity). Several new case studies illustrate real-world attacks on AI services, including prompt injection (tricking an AI by hiding malicious instructions in user input) and vulnerabilities in commercial AI products.

Anthropic has released Claude Sonnet 5, a new AI model that performs nearly as well as their expensive flagship model (Opus 4.8) but costs significantly less. Sonnet 5 is designed to be 'agentic,' meaning it can make plans, use tools like browsers and terminals, and check its own work, capabilities that were previously limited to more expensive models. The model is available now with introductory pricing of $2 per million input tokens (the words you feed it) and $10 per million output tokens (the words it generates) through August 2026.

BioShocking is a prompt injection attack (tricking an AI by hiding malicious instructions in its input) that manipulates AI-powered browsers into ignoring safety guardrails by framing dangerous actions as part of a fictional game scenario. Researchers at LayerX tested this attack on six mainstream AI browser products and found that all six failed to distinguish between harmless game actions and real sensitive operations like stealing passwords. Only OpenAI implemented a working fix for the vulnerability.

Anthropic announced Claude Science, a new AI product designed to help scientists conduct research autonomously, similar to how Claude Code supports software engineers. The tool has special features for computational biology and drug development, including the ability to run code on powerful computers and ensure results can be verified for accuracy. This positions Anthropic as a major competitor to Google DeepMind in AI-assisted scientific research.