Frontier-Capable Adversaries Outside. Unchecked AI Agents Inside. Now What?

Two shifts are hitting security programs at once. Frontier AI has collapsed the cost and time of attack, and open-source models are close behind. At the same time, AI agents are going live inside the business applications you already run, holding credentials and acting on their own. Most of the advice on offer is either panic or a product pitch. The uncomfortable truth is that guardrails are not reliably holding on either front.

Avivah Litan created and led Gartner’s AI Trust, Risk and Security Management research during more than two decades as a Distinguished VP Analyst. She will argue that today’s risk-mitigation methods need updating and more agility. Her talk starts with a pragmatic approach to where to begin when adversarial capability surges.

Then she takes on the harder questions: • Your blind spots: agents are switched on inside your organization’s applications, often by default and outside security’s view. Each one holds credentials, reads untrusted data, and can act. Computer-use agents and sprawling shadow AI make it worse. • Guardrails are not controls: models can ignore instructions and talk their way past their own safeguards. What happens when the thing holding the credential refuses your rules, is persuaded to, or simply acts against your interests? • Delegated authority: when an agent can plan and act on its own, which existing controls still matter, and where do you need new ones?

Shweta Lakshmanan of Obsidian Security’s product team will then show what it takes to act. Against frontier-capable attackers, the answer isn’t more findings but more leverage: knowing which few fixes take away the most attacker options. Against agent risk, the answer isn’t more visibility alone but enforcement wherever agents run, with enough business context to tell legitimate actions from dangerous ones. Shweta will show what it takes to get there. The session closes with open Q&A.