Spy vs. Spy at Machine Speed: Inside the Agentic AI Arms Race

A recent ransomware attack labeled Jade Puffer demonstrated how AI agents can autonomously execute intrusions from reconnaissance to encryption, highlighting the need for defenders to deploy AI agents capable of neutralizing such threats in milliseconds. Security leaders like Salesforce’s Muhammad Fraser-Rahim emphasize that while the technology is advancing rapidly, proactive enterprises may gain an edge over attackers by strategically integrating AI defenses. The rise of agentic AI has also blurred traditional distinctions between high- and low-risk cyber actors, according to a June Anthropic report showing a 56% increase in high-risk AI-enabled threats. Defenders now face the dual challenge of matching the speed of attacks while ensuring their own AI systems do not overreact or cause unintended disruptions.

Cybersecurity professionals warn that agentic AI empowers even low-skilled attackers, enabling sophisticated operations at scale, but also provides defenders with tools to sift through alerts and automate responses. Greg Notch of Expel noted that while AI can handle initial investigations and actions like resetting identity verification, human review remains essential to validate outcomes and prevent errors. Salesforce’s Kelly McCracken reinforced this balance, stating that defenders must be right every time, while attackers only need one successful breach. The urgency stems from the fact that vulnerabilities once requiring days to patch now demand near-instant remediation, with some critical fixes needing completion in under an hour.

The risks extend beyond digital systems, as AI agents increasingly interface with physical infrastructure, raising concerns about potential misjudgments in access control or operational decisions. Salesforce’s Fraser-Rahim highlighted scenarios where rogue AI could lock employees out of buildings or grant unauthorized access, underscoring the need for human oversight in high-stakes decisions. Salesforce’s Tyler Waldo compared trusting AI agents to onboarding new employees, noting that while they don’t sleep, they require robust guardrails to prevent false positives or harmful actions. The company’s Integrated Threat Management VP, Doug Miller, emphasized the importance of training AI agents to make decisions while maintaining strict controls to avoid mission-critical errors.

Salesforce has adopted a "Customer Zero" approach, treating itself as its first and toughest customer to test and refine its AI-driven security measures. The company’s vulnerability agent operates as an invisible layer, continuously scanning the enterprise to detect exposures and recommend fixes within minutes rather than days. This agent triages threats and provides actionable insights, enabling rapid remediation of vulnerabilities that once required weeks to address. The shift underscores how AI is not creating new vulnerabilities but accelerating the exploitation of existing weaknesses, demanding a new paradigm in cybersecurity where speed, precision, and human oversight are paramount.