← Back to articles
AI SecurityMembers
AI Summit at Black Hat: Red Teaming LLMs on the Main Stage
By S. Chen · June 5, 2026 · 10 min
Highlights from the AI Summit — where prompt injection took a back seat to agent tool misuse and model supply-chain attacks.
The moment you hand an agent a shell, a browser, and a credit card, your threat model looks a lot like insider risk.
Scope every tool. Rate-limit every tool. Log every tool call with the exact arguments and the exact response.
Assume the model will be convinced to do the wrong thing eventually. Design so 'the wrong thing' is survivable.