OpenAPA brings deterministic prompt-injection defense to AI agents

OpenAPA applies a military-style data classification model to AI agents: a session inherits the highest classification level of the data it reads, and any attempt to write to a lower-trust destination is blocked by policy rather than by prompting. The approach is prompt-independent but lowers end-to-end completion rates and raises token costs.

• Agent session inherits the highest classification level of data it reads
• Writes to lower-trust destinations are blocked deterministically
• Defense is prompt-independent and resists prompt injection
• Trade-offs: lower completion rate and higher token usage