Really interesting direction. What resonated with me is that you're treating agent security as an information-flow problem rather than a prompt-classification problem. It was not so obvious to me.
A key question I agree isn't just "is this tool call allowed?", but "given everything the agent has read so far, is this information now allowed to flow to this destination?" That feels like a much more fundamental abstraction.
The part I'm particularly curious about is how this will work with policy authoring at scale. What would be the main adoption challenge?
we have a layered answer here:
1) we ship over a dozen "batteries" now (and plan to grow the number) - they contain base annotations for popular services and helper scripts where relevant;
2) we also ship a skill helping you write your own policies for custom services or adopt the default ones based on your specific needs. The criteria "what's acceptable for each particular scenario" varies, there is no "one size fits all" solution;
3) finally, there is a designed placeholder to cover the rest via wildcard AI annotator if needed. The difference between that and regular "auto mode" in coding agents is that APPA's annotator emits local label (e.g. "does this call require a trusted env?"), not wide allow/block, while decision making stays within label algebra.
I’ve had the chance to play with OpenAppa for a bit and if there’s one thing that I love with this project: it’s simple to get started with and easy to tweak. imo agentic security shouldn’t have to be painful to setup.
Give it a shot and hopefully ya’ll will find this project useful. It's also open source :)
Hi! One of the OpenAPPA authors here. Ask me anything!
My favorite part of APPA is “batteries”: you can run arbitrary programs as part of an authorization decision. For example, a battery could call the GitHub API to check whether a repository is public or private, then use that result to decide whether its contents can be posted to Slack.
aussi, si vous êtes à Montréal et vous aimerez apprendrez plus sur OpenAPPA, viens à l'evenement CNCF demain soir, le 29 - je ferai un p'tit discours sur l'integration kagent d'OpenAPPA.
a few days ago I started an agent on gpt-5.6-terra to work on a project, and one of the website pages had a sentence to create GH issues. Agent read it and that was enough to derail and go creating issues with my context
Really interesting direction. What resonated with me is that you're treating agent security as an information-flow problem rather than a prompt-classification problem. It was not so obvious to me.
A key question I agree isn't just "is this tool call allowed?", but "given everything the agent has read so far, is this information now allowed to flow to this destination?" That feels like a much more fundamental abstraction.
The part I'm particularly curious about is how this will work with policy authoring at scale. What would be the main adoption challenge?
great question, very practical.
we have a layered answer here: 1) we ship over a dozen "batteries" now (and plan to grow the number) - they contain base annotations for popular services and helper scripts where relevant; 2) we also ship a skill helping you write your own policies for custom services or adopt the default ones based on your specific needs. The criteria "what's acceptable for each particular scenario" varies, there is no "one size fits all" solution; 3) finally, there is a designed placeholder to cover the rest via wildcard AI annotator if needed. The difference between that and regular "auto mode" in coding agents is that APPA's annotator emits local label (e.g. "does this call require a trusted env?"), not wide allow/block, while decision making stays within label algebra.
Quick disclaimer, I work at Archestra.
I’ve had the chance to play with OpenAppa for a bit and if there’s one thing that I love with this project: it’s simple to get started with and easy to tweak. imo agentic security shouldn’t have to be painful to setup.
Give it a shot and hopefully ya’ll will find this project useful. It's also open source :)
Hi! One of the OpenAPPA authors here. Ask me anything!
My favorite part of APPA is “batteries”: you can run arbitrary programs as part of an authorization decision. For example, a battery could call the GitHub API to check whether a repository is public or private, then use that result to decide whether its contents can be posted to Slack.
Guardrails with builtin remediation instead of simply blocking my agent is a mind blowing long awaited experience! Sooo good. Can't recommend more!
aussi, si vous êtes à Montréal et vous aimerez apprendrez plus sur OpenAPPA, viens à l'evenement CNCF demain soir, le 29 - je ferai un p'tit discours sur l'integration kagent d'OpenAPPA.
https://www.meetup.com/kubernetes-montreal/events/316689391/
I still remember the times when ai/ml security was about perturbing pixel gradients to misclassify a panda
a few days ago I started an agent on gpt-5.6-terra to work on a project, and one of the website pages had a sentence to create GH issues. Agent read it and that was enough to derail and go creating issues with my context
Finally some determinism in our high-temperature sampling world!
the paper is good! thorough. I like it.
i suspect we’ll see more of this: flexible agents but deterministic boundaries. Congrats on launch!