Who’s watching the AI watchers?

A non-profit watchdog, Protect Democracy, is suing four federal agencies over their opaque process for reviewing frontier AI models before release. The lawsuit aims to uncover details about evaluation frameworks, participant companies, and oversight authority, arguing that secrecy impedes transparency and accountability. This issue emerges amidst discussions in Congress on liability protections under the Cybersecurity Information Sharing Act, which depend on understanding government review processes.

Meter, a nonprofit that tests AI models, disclosed two cyberattacks from earlier in the year. In one incident, attackers exploited a vulnerability to access API keys, resulting in significant cloud credits being consumed over weeks. In another, attackers conducted probing campaigns using automated tools, though no sensitive data was confirmed as accessed, and some vulnerable evaluation data was exposed without evidence of exploitation. Separately, leaked documents unveiled a Russian military cyber training pipeline at Bauman Moscow State Technical University, revealing extensive programs preparing students for cyber espionage and influence operations, linked to Russian military units and threat groups.

Huntress uncovered a campaign where attackers used rogue ScreenConnect installations and social engineering to spread malware across multiple organizations. The malware employed a four-stage script chain to profile infected systems, establish persistence, and deploy various malicious tools like cryptocurrency miners and backdoors. A notable feature was a modified ScreenConnect client that automatically infected connected machines, facilitating worm-like spread. Users are advised to rebuild compromised systems and review audit logs.

Thomson Reuters revealed a breach in its court case management system, C-Track, exposing appellate court records across numerous US jurisdictions and Ontario, Canada. The breach occurred in March, but detection was delayed until June. Sensitive information such as personal identifiers and legal documents may have been exposed, with affected individuals offered credit monitoring. The breach highlights ongoing risks in legal data security.

Palo Alto Networks' Unit 42 identified a voice phishing campaign, Spring Ring, where attackers impersonated IT support on Microsoft Teams to deceive employees into installing malware and gaining system access. The attack exploited trust in collaboration tools rather than vulnerabilities. Meanwhile, Plex urged users to patch security flaws in its media software, and Cisco warned of unpatched vulnerabilities in its email encryption products and network hardware, emphasizing the need for timely updates to prevent exploitation.

Rob van der Veer, Chief AI Officer at Software Improvement Group, discussed the growing AI attack surface, focusing on shadow AI, open source dependencies, and AI-generated code. He emphasized the importance of managing vulnerabilities in dependencies, treating shadow AI use as a demand signal for in-house AI provision, and developing frameworks for AI code review. He also stressed the need for organizations to foster collaboration, share best practices, and balance AI tools with human oversight to mitigate security risks. The talk highlighted the challenge of keeping security measures ahead of rapidly evolving AI technologies.

The FCC proposed a public scorecard to help consumers evaluate their phone providers' effectiveness in preventing robocalls, based on real-world outcomes like complaints and enforcement actions. The agency has already taken action by removing non-compliant providers from its database. The discussion pointed to ongoing efforts to both incentivize and enforce better telecom security, with potential sanctions against persistent robocallers. The overall message was that transparency and accountability are key to improving consumer trust in telecom services.