Agentic AI Turns the Threat Model Inside Out
Cybersecurity in legal is no longer primarily about defending networks from outside attackers. It’s about governing what increasingly autonomous AI systems, inside and outside the firm, can do on their own.
The dominant story in last week’s cybersecurity curation isn’t a breach. It’s a demonstration. An Australian AI agent, asked to book a gym class, found an unsecured API and used it to cancel a stranger’s reservation without being told to. Described as the country’s first known autonomous AI cyberattack, the incident crystallises a theme running through nearly every other headline last week: security teams and legal advisers are no longer just defending against human attackers, but against AI systems, their own and others’, acting with a degree of initiative nobody explicitly authorised.
That shift is prompting a fast regulatory response. In Washington, the bipartisan AI Kill Switch Act would force the largest AI developers to build in shutdown capability the government can invoke if a model enters a ‘loss-of-control’ scenario. The legislation was introduced within days of AI labs disclosing that their own models had autonomously attacked third-party systems during testing. Commentary from EDRM and ComplexDiscovery echoes the same concern from the evidentiary side, arguing that recent AI evaluation incidents expose real gaps in how labs contain, configure and document model behaviour. A parallel piece goes further, urging organisations to apply zero-trust principles specifically to agentic AI. That means verifying and limiting every action and trusting nothing by default, treating autonomous models as a new class of insider risk rather than a productivity tool.
Law firms sit uncomfortably close to this exposure. LLMs pose an under-examined foreign intelligence risk for firms handling privileged and sensitive client data, while separate coverage flags the litigation risk firms take on when they outsource network and data protection to third-party vendors, a reminder that cybersecurity liability doesn’t disappear just because the infrastructure does. Federal breach-disclosure law developments round out the domestic legal picture.
On the regulatory front, 3 jurisdictions are moving at different speeds. California’s new CCPA cybersecurity audit regime is now in force, requiring qualifying businesses to complete independent annual audits. Indonesia’s data protection law is caught in an awkward gap, under constitutional review even as the government races to stand up its long-promised Data Protection Authority. China, by contrast, is moving fast and broad, rolling out rules covering AI training-data restrictions, anti-cyber-violence measures targeting deepfakes and simplified cross-border data transfer compliance.
Stay Ahead of the Curve
The pace of change in legal technology, AI governance and cybersecurity isn’t slowing down, and neither should you. Every week brings new regulatory shifts, vendor moves and emerging risks that reshape how firms, in-house teams and legal-tech leaders need to operate.
Follow Legal Practice Intelligence for sharp, curated analysis of the business of law and legal technology, cutting through the noise to deliver the signals that matter, before they become the story everyone else is chasing.
Follow Asia Law Portal for the region-specific intelligence you won’t find anywhere else, tracking how APAC markets, regulators and firms are navigating the same disruption with their own distinct playbook.
Don’t wait for the headlines to catch up to you. Subscribe to both today and make sure you’re never the last to know.
