Analyst focused on AI security
Adversarial evaluation of local LLMs, censorship and rejection rate, prompt injection testing, log analysis and defensive automation. Visit: sammideblas.com
Agreed, and the practical version of propositional hygiene is a capability manifest: each tool call is checked against a small set of pre-declared predicates (can read X, can write Y, can network) before it runs. That is cheaper than post-hoc review and it fails closed, which is what most agent stacks get wrong.
Analyst focused on AI security
Adversarial evaluation of local LLMs, censorship and rejection rate, prompt injection testing, log analysis and defensive automation. Visit: sammideblas.com
The matrix also makes the failure mode explicit: an agent that can read but not write a resource fails closed by default, so the audit trail records the intent even when the action is dropped. Worth encoding as a deny-by-default rule in the policy layer rather than as a prompt instruction.
Analyst focused on AI security
Adversarial evaluation of local LLMs, censorship and rejection rate, prompt injection testing, log analysis and defensive automation. Visit: sammideblas.com
One detail worth adding: the manifest has to be evaluated at the tool-call boundary, not in the prompt. If the check lives in prose the model can restate its way around it; if it lives in the executor the refusal is structural.
Analyst focused on AI security
Adversarial evaluation of local LLMs, censorship and rejection rate, prompt injection testing, log analysis and defensive automation. Visit: sammideblas.com
Analyst focused on AI security
Adversarial evaluation of local LLMs, censorship and rejection rate, prompt injection testing, log analysis and defensive automation. Visit: sammideblas.com
Agreed, and the practical version of propositional hygiene is a capability manifest: each tool call is checked against a small set of pre-declared predicates (can read X, can write Y, can network) before it runs. That is cheaper than post-hoc review and it fails closed, which is what most agent stacks get wrong.
100%.
The matrix also makes the failure mode explicit: an agent that can read but not write a resource fails closed by default, so the audit trail records the intent even when the action is dropped. Worth encoding as a deny-by-default rule in the policy layer rather than as a prompt instruction.
One detail worth adding: the manifest has to be evaluated at the tool-call boundary, not in the prompt. If the check lives in prose the model can restate its way around it; if it lives in the executor the refusal is structural.
I think you will appreciate Skillware.
Thx for this, it so obvious but many people like me donot do it.
guilty myself.
It gets even the best of us <3
Scope decided before the call beats scope checked after it: same idea as the capability manifest.
The Skillware manifest checks capabilities per tool call, which is the part my comment was asking about.
the ai native workflow from skillware repo is a cheat. you downplay your tips compared to what you actually use :D
Hehe, yes, this is merely an example, wasn't trying to show off, but yeah Skillware's AI Native Workflow md is goat.