Strategy
An alliance of more than 100 technology companies and other organizations has proposed establishing the AI agent safety reporting system SAFE to share security incident information, identify common risks, and issue defense recommendations. The system was developed by the Open Security AI Alliance working group, and the Linux Foundation has released a draft for public comment.
A report released by CrowdStrike on Monday noted that malicious use of AI has rapidly accelerated, with some threat groups embedding AI in their cyber operations. China-linked groups Vault Panda and Genesis Panda used AI to exploit critical vulnerabilities within 24 hours of proof-of-concept disclosure; North Korea-linked group Stardust Chollima injected malicious npm packages into 131 trusted Mastra AI frameworks through supply chain attacks. CrowdStrike's head of counter-adversary operations, Adam Meyers, warned that the vulnerability exploitation window has shortened to hours, with zero-day vulnerability weaponization outpacing traditional patch cycles.
U.S. water facilities were attacked by hackers suspected of being linked to Iran, affecting at least seven states. The OT Cybersecurity Alliance calls on CISA to issue binding operational directives and asks Congress to take three specific measures to strengthen critical infrastructure security.
A Unit 42 report reveals that a Chinese-speaking attacker used DeepSeek and the Hermes Agent framework for autonomous cyberattacks, scanning GitHub for popular PoCs and attempting to exploit vulnerabilities in Langflow, n8n, and others, but most attacks failed due to authentication requirements, with only manual exploitation of Citrix NetScaler and Marimo vulnerabilities succeeding.
A report released by Okta on Wednesday indicated that 81% of CISOs believe their AI systems lack proper governance, and only 47% of enterprises are aware of all AI agents on their networks. Shadow AI is widespread, leadership alignment is insufficient, and concerns among U.S. CISOs about AI-driven attacks are particularly pronounced.
IBM's latest report indicates that the average cost of a data breach in 2025 reached $5 million, a 12% year-over-year increase. Although half of breached organizations used AI agents for threat hunting, fewer than one-fifth used them for vulnerability management. Attacks on AI models are more costly, with 92% of victim organizations failing to effectively control AI access, and shadow AI-related incidents doubling, highlighting governance lag.
Microsoft announced on Monday the launch of Project Perception, an agentic security system designed to help defenders counter new types of attacks where malicious actors leverage AI. The company also released the AI model MAI-Cyber-1-Flash, integrated with the MDASH code scanning tool to identify and fix vulnerabilities. This move is part of an industry push to advance AI security technologies to protect enterprise IT and critical infrastructure.
Security firm Arctic Wolf's annual report, released on Tuesday, indicates that businesses are more worried about AI-driven threats than traditional cybersecurity risks, which may weaken their preparedness for the attacks they are most likely to face. About one-third of organizations listed AI as their top cybersecurity concern, while worries about malware, credential theft, and cloud misconfigurations declined compared to 2025. The report also revealed that 63% of organizations experienced at least one cybersecurity incident in the past 12 months, yet leader confidence remains high.
Global technology, cybersecurity, and AI companies are intensifying efforts to push the US government to embrace open AI systems, highlighting their safety benefits and warning against obstructing them in the US-China AI competition. On Friday, 76 companies issued an open letter praising open-weight models, and on Monday, 37 enterprises formed the Open Safety AI Alliance to advance a similar mission.
The Trojan horse story is often used as a metaphor for AI safety risks, but this article argues that Agentic AI does not create a new category of risk, but rather amplifies existing patterns of exploiting trust in supply chain attacks. Drawing on discussions from Cympire and Forbes, the article compares traditional Trojans with AI agents, proposing a return to third-party risk management frameworks, centering on 'permissions' rather than 'access,' and using four key questions to construct a governance strategy.