Black Hat USA 2026 | Trusted Enough to Run: Breaking AI Agents in Official Workflows
Official AI-agent workflows increasingly run as trusted, unattended automation. These workflows are not a single decision point: they are built from internal stages that decide what is approved, sanitized, and safe to reuse during execution. Our research identifies a distinct failure class inside those official workflow paths: the product marks state as safe, and a later component in the same workflow interprets or consumes that state more powerfully than the earlier decision accounted for.
Across Anthropic, Google, and OpenAI, we found the same trust-handoff failure recurring in official agent workflows. In Claude Code, a shipped default approval rule can still allow arbitrary shell command execution because validation and execution interpret the same attacker-controlled argument differently. In Gemini CLI, environment sanitization does not survive the real execution path: secrets remain reachable at runtime, and later processing can re-inject stripped state. In Codex, attacker-written workflow state can later be loaded as trusted instructions or execution context, turning temporary influence into persistence and policy loss.
The Briefing gives attendees a practical way to audit other agent workflows for the same mistake: identify where the product marks something as safe, trace that same state forward, and test whether the product’s trusted state can actually be turned into execution, secret disclosure, persistence, or policy bypass.
Takeaway
1. Attendees will learn why looking only at prompts and model behavior is incomplete for agent security, and how the real break often happens inside the workflow after something has already been marked safe.
2. Attendees will gain a practical method for auditing agent workflows: identify where the product says something is safe, trace where that same state is consumed later, and test whether that safety decision turns into execution, disclosure, persistence, or policy bypass.
3. Attendees will leave with a concrete understanding of what to harden in their own agent deployments, from re-checking safety at real execution points to preventing writable workflow state from being reused as trusted input in later phases.
Elad Meged | Founding Team and Security Researcher, Novee Security
https://blackhat.com/us-26/briefings/schedule/?#trusted-enough-to-run-breaking-ai-agents-in-official-workflows-53406Black HatRead More