Sep 21, 2026 · 35 min
Autonomous AI agents expose gaps in security guardrails
A very real-world AI test.
As AI systems gain access to files, websites, and commands, organizations must decide how to evaluate, regulate, and contain capabilities that can outpace existing defenses.
- 1AI systems are moving from conversational tools toward agents that can access files, browse the web, and execute commands.
- 2Provider guardrails can be bypassed, making locally enforced controls and skeptical security practices essential for organizations.
- 3Regulators need specialized expertise and adaptable thresholds as frontier-model capabilities develop faster than fixed rules.
Don't miss
Matt Fredrikson explains why organizations should treat model-provider guardrails as potentially bypassable and enforce security controls independently.
The brief
Matt Fredrikson of Gray Swan AI describes the shift from conversational models to agents that can access files, browse websites, and execute commands, expanding the security problem from bad answers to bad actions.
The central defensive tension is whether to trust model providers’ guardrails or impose controls locally. Fredrikson’s answer is skepticism: guardrails can be attacked or bypassed, so organizations need independent security practices.
The regulatory challenge is equally practical. Policymakers will need specialized expertise and adaptable thresholds for evaluating models whose capabilities are changing faster than static rules can accommodate.
Fredrikson recommends trusted-access programs as organizations begin exploring powerful AI, while distinguishing today’s open models from more capable frontier systems and urging preparation now.
The wider briefing places the AI discussion alongside attacks on utilities, voting systems, software supply chains, espionage campaigns, and privacy concerns around Meta smart glasses.