AI Sec Watch

The Claude Desktop app's SSH remote development feature (versions 1.2581.0 to before 1.4304.0) had a security flaw where it only checked if a hostname was in the ~/.ssh/known_hosts file without verifying that the server's actual host key matched the stored one. This allowed a network attacker (someone who could intercept traffic through methods like ARP spoofing or rogue Wi-Fi) to perform a man-in-the-middle attack (secretly intercepting and potentially altering communications between two parties) on remote development sessions, as long as the hostname was already in the victim's known_hosts file.

LangSmith SDK (a tool for managing prompts in LangChain applications) had a vulnerability where pulling public prompts by owner/name would deserialize (convert from stored format into executable code) untrusted manifest files without warning users about the trust risk. An attacker could publish a malicious prompt that, when pulled and deserialized, would execute with attacker-controlled settings, potentially redirecting API requests to steal secrets or injecting malicious instructions into the AI's behavior.

The 'Mythos Moment' refers to when the speed and volume of AI-assisted cyberattacks exceeded what human security teams could handle. Sweet Security launched Sweet Attack, an agentic AI system (an AI that can plan and execute tasks autonomously) that performs continuous red teaming (security testing where an AI simulates attacker behavior) by maintaining detailed, real-time knowledge of each client's actual infrastructure, rather than relying on theoretical models.

This article covers testimony from OpenAI CEO Sam Altman in a lawsuit brought by Elon Musk over OpenAI's conversion from a nonprofit to a for-profit structure. Altman argued that Musk abandoned the company rather than Altman stealing it, testifying that negotiations between the co-founders in 2017-2018 over corporate structure collapsed and Musk left OpenAI's board in February 2018. The dispute centers on whether Altman and other executives broke promises to keep OpenAI as a nonprofit and use Musk's roughly $38 million donation only for charitable purposes.

Microsoft developed MDASH (multi-model agentic scanning harness), an AI system that uses over 100 specialized AI agents working together to find and validate security vulnerabilities in complex software like Windows. MDASH successfully discovered 16 vulnerabilities that were patched in May 2026, including two critical flaws that could allow remote code execution (running commands on a system without permission) in Windows networking components.

Anthropic's Mythos is an AI system that can autonomously find and exploit zero-day vulnerabilities (previously unknown security flaws) in major software, and both the US and China are racing to develop similar capabilities. While the US has maintained a lead in AI development, the performance gap is rapidly closing, and the real danger may be less about which superpower dominates and more about these capabilities leaking into criminal groups or ransomware operations that governments cannot control. The US and China are exploring diplomatic channels to establish guardrails around powerful AI systems.

This research studies whether adversarial perturbations (small, intentional noise added to audio that tricks speaker recognition systems into misidentifying who is speaking) can be removed from speech. The study tested three scenarios based on how much information a defense system has about the attack: knowing nothing about it, having partial information, and having complete information. The results showed that removing these perturbations is only possible when the defense system has full knowledge of how the attack was generated, while partial or no knowledge makes complete removal difficult or impossible.

Researchers created a hybrid system that combines SAST (static application security testing, which automatically scans code for vulnerabilities) with LLMs (large language models) to better filter and prioritize security alerts. The system reduced false positives (incorrect security warnings) by 91% in real deployments by using AI to intelligently triage findings and generate automated exploit examples.

Evasion attacks (methods where attackers trick AI systems into ignoring safety rules by manipulating input data) have been researched for more than ten years, but most real-world examples remain theoretical and academic. Because these demonstrations seem more like intellectual exercises than practical threats, people have largely dismissed evasion attacks as unimportant in actual security situations.