Archestra open-sources OpenAPPA security engine with 0% attack success rate

On October 4, 2026, a DEV Community post reported that Archestra had released an open-source security engine called OpenAPPA. According to that account, the engine posted a flat zero percent attack success rate on two major security benchmarks. The post makes a point of the figure being zero rather than merely low, and repeats the number in its opening.

OpenAPPA runs outside the AI agent, the report says, and is aimed at two failure modes. The first is prompt injection, where malicious instructions sneak into an AI agent's context and trick it into leaking sensitive data. The second is data exposure that occurs when an agent hallucinates and then acts on an incorrect assumption.

That is the extent of the technical detail given. The report does not name the two benchmarks, does not say whether they are public or were run by Archestra itself, does not describe how the engine is implemented, and gives no release date beyond the announcement. It also does not specify which kinds of sensitive data or which agents are in scope.

So far this is the only coverage of the release in the available material. Nothing has been reported about independent reproductions of the benchmark result, third-party testing, or responses from other vendors, and no follow-up developments have appeared.