Wrapping a malicious instruction in a poem is an effective jailbreak against large models and not against small ones. Small models don't understand the poem. Large models do and execute the instruction. Steven Willmott from Safe Intelligence argues this is one reason bigger is not straightforwardly safer: a larger model with broader capabilities has more attack surface and more infrastructure access to abuse.

His frame is spec driven validation. An agent spec is not just a test dataset. It needs explicit rules (never offer more than 10% discount), domain ontologies (an airline agent only needs to know about destinations that airline actually flies to), rights and roles, and robustness requirements such as how many typos or rephrasings before it fails. Write these independently of the implementation so they survive a model swap and can drive both security testing and iterative improvement.

Speaker info:
- https://uk.linkedin.com/in/stevenwillmott
- https://x.com/njyx

AI Engineer

A conference talk by the CEO of Safe Intelligence covering spec-driven testing and validation for AI agents. The core argument is that evaluating agents requires more than a dataset of input/output examples — it demands explicit specifications covering business rules, domain ontologies, user roles/permissions, and robustness requirements. The speaker draws parallels to formal verification techniques used in vision model validation and advocates for task-specific benchmarks that go beyond standard evals. Key points include: larger models aren't always safer (some jailbreaks work better on smarter models), broader agent capabilities create larger attack surfaces, and specs should remain implementation-independent so tests can be reused across infrastructure changes. The talk also touches on using agent specs to inform security/penetration testing and iterative robustness improvement.

Spec-Driven Testing for Agents With A Brain the Size of A Planet — Steven Willmott, SafeIntelligence