Most AI governance stops at the dashboard. But the gaps that actually bite in production are the ones between the layers — shadow AI, agent delegation, prompt-level leaks, expiry, and unprovable audit trails. A field guide
Most "AI governance" I've seen — and I've been building in this space — converges on the same shape: a dashboard, a log, a policy page. That's necessary. It's also where the thinking usually stops, and the gaps that actually cause incidents live between those pieces, not inside them.
This is a field guide to five blind spots I keep running into. None of them are exotic. They're the boring, structural gaps that don't show up until an agent does something nobody authorized and you're trying to reconstruct why. For each one I'll describe the problem, why the common approach misses it, and the shape of a fix. I'll mention what I've built where it's relevant, but the point here is the problem, not a product — most of these you can address with tools you already have.
Blind spot 1: Shadow AI is a network problem, not a policy problem
Every company has a policy about AI use. Almost none can tell you what AI is actually running. Someone spun up a local Ollama instance. A team wired an unsanctioned API into a cron job. A browser extension is quietly shipping prompts somewhere.
Why the common approach misses it: policies govern sanctioned usage — the tools you already know about. Shadow AI is by definition the stuff you don't. You can't write a policy for a model you can't see.
The shape of a fix: treat discovery as a detection problem, the way you'd treat rogue devices on a network. Endpoint signals (processes like ollama/vllm, listening ports, model-weight files on disk) plus passive network discovery surface what's actually running, independent of anyone declaring it. The key design rule: discovery must be read-only. Finding a service is not the same as connecting to it — nothing should auto-attach to a discovered endpoint without an explicit human decision, or your discovery tool becomes its own attack surface.
Blind spot 2: Prompt-level leaks happen before your logs ever see them
You log the request. Good. But by the time it's in your log, the secret has already left — the API key, the customer PII, the internal hostname — pasted into a prompt and shipped to a third-party provider.
Why the common approach misses it: observability is downstream. It records what happened; it doesn't intervene. Logging a leak is not preventing a leak.
The shape of a fix: put a masking chokepoint before the prompt leaves. Scrub secrets (cards, keys, IDs, PII) at the gateway, so the provider never receives raw sensitive input in the first place. A subtle upside: gateway-level masking protects you from the provider itself, not just from your own traces. App-level redaction only protects your logs; the prompt still leaves your building intact. (This one bites hardest with LLM providers — the data's gone the moment the request fires.)
Blind spot 3: Agent delegation has no subset guarantee
Single agents were easy to reason about. Now an orchestrator calls a researcher, which calls a writer, which calls a tool. Each hop hands some authority to the next — and in most setups, nothing structurally prevents a sub-agent from using a capability it was never granted.
Why the common approach misses it: prompt instructions ("you may only read, never write") get reasoned around. A capability boundary enforced in text is a suggestion. Under enough pressure — a retry, a clever completion, a tool that reframes "delete" as "cleanup" — it leaks.
The shape of a fix: enforce capability subsetting at every hop, structurally. A sub-agent can never hold more than the agent that spawned it. If the parent can't deploy, nothing it spawns can deploy, regardless of how the chain grows. This turns a prompt rule into an invariant. And log every rejected escalation separately with the exact over-requested capability — the rejections are the signal, because the attacker or bug doesn't know which check failed.
Blind spot 4: Delegations don't expire, so revocation is a fantasy
Say you do scope a delegation correctly. A parent hands a sub-agent a credential for a three-minute task. Then the sub-agent spawns an async job, or retries against a queue, and the work outlives the authorization. Now that credential is either expired mid-flight, or scoped so wide in duration that "revoking" it means manually hunting it down.
Why the common approach misses it: capability subsetting handles what, not for how long. Time is the dimension everyone forgets, and it behaves differently — capabilities only ever shrink down a chain, but time, once you introduce async work, doesn't stay neat.
The shape of a fix: monotonic TTL — a delegation can't outlive the one that authorized it, the temporal twin of the subset check. Enforce it at delegation time (a longer TTL is a violation) and again at action time (an action under an expired delegation is denied). The async case is genuinely hard and I don't think it's fully solved — the cleanest direction I've found is treating a queued job as a new delegation request at execution time rather than reusing the original grant, but that just relocates the question to "who signs the re-delegation."
Blind spot 5: A log you control is not proof
This is the one that reframes all the others. When an auditor — or a regulator under something like the EU AI Act — asks "who authorized this agent to do that?", you point at your audit trail. But your audit trail is a claim. If your server wrote the log, your server could have written anything. Monitoring tells you what a system says happened. It doesn't let anyone verify it independently.
Why the common approach misses it: the entire industry equates "observable" with "accountable." They're not the same. A dashboard is evidence to you. It is not evidence to someone who doesn't trust you.
The shape of a fix: make the authorization itself cryptographically verifiable. Sign each delegation (Ed25519 is a good fit — small keys, fast verification, deterministic). Store the exact signed payload so nothing has to be reconstructed. Then anyone can verify the signature against the signer's public key in their own browser, without trusting the server — the server holds only the public key, so it can verify but never forge. "Trust me, here's my log" becomes "here's the math, check it yourself." Put the delegation's expiry inside the signed payload too, so a lifetime can't be extended after the fact.
The through-line
Notice that four of the five aren't about any single system being wrong. They're about the seams — between sanctioned and shadow, between logging and intervening, between one agent and the next, between authorization and time. Governance tools tend to instrument the boxes and ignore the lines connecting them, which is exactly where autonomy leaks.
I've been building an open-source, self-hosted platform (Provenza) that tries to close these seams — the name is from provenance, because the thread running through all five is making the origin of an AI action provable rather than merely observed. It's early and I'm figuring plenty out, especially the async-expiry case above. But you don't need my project to take the ideas: discovery as detection, masking before the prompt leaves, subset-at-every-hop, monotonic TTL, and signed authorizations are all things you can build into what you already run.
If you're working on agent infrastructure, I'd genuinely like to hear which of these you've hit — and how you're handling the ones I clearly haven't fully solved. The comments on my last posts
thought me more than the posts ded.
Thanks to Pushpendra Agrawal and Reid Marlow — their comments on my earlier posts directly shaped blind spots #3 and #4 here. The rejected-escalation logging and the monotonic-TTL framing both came out of those conversations. This is the kind of feedback that turns a post into a feature.
GitHub: https://github.com/kironovlaziz-del/provenza
git clone: https://github.com/kironovlaziz-del/provenza.git
Top comments (0)