Research review / trendwatch

Trend Watch: An Agent Control Only Sees What Crosses It

This week's evidence keeps returning to placement. A framework measures which of 15 agent action paths a runtime control can see, a shared GPU leaks which approval setting an agent runs under, an LLM judge passes forbidden database writes that a simple rule check catches, and default tracing exported OAuth tokens. Where a control or record sits decides what it can see and what it exposes.

By Sandeep Belgavi · Published 2026-11-14

Most agent safety discussions ask whether a check is smart enough. This week's papers and fixes ask a different question: is the check in the right place? A gateway, a judge or a trace can only see what passes through it, and can expose whatever does.

Four signals

  • Gateways miss side doors. guardscope (November 11) starts from the point that "a control can only evaluate the actions that cross the point where it is placed, and an agent can act through more paths than the tool-call channel." It predicts which of an agent's "15 action paths" a control can check, then classifies each executed action as "checked, bypassed, or prevented" against kernel-level ground truth, using real controls including Microsoft's agent governance toolkit and agentgateway.
  • Configuration leaks through hardware. PolicySniff (November 12) shows a co-resident process timing its own GPU kernels can infer whether an agent is set to ask before consequential tool calls, at "94.17% balanced accuracy" on a sparse mixture-of-experts runtime. "Behavioral compliance does not imply configuration confidentiality."
  • Judges read transcripts, not effects. In When Evaluators of AI Agents Disagree (November 10), across 2,160 tau2-bench episodes, agents made a forbidden write in 11. "The judge passed 10 to 11, with or without the policy or transcript. A stronger judge still passed 6." A rule-based check of expected against actual writes, "with no model call, caught all 11".
  • Records expose what they capture. ADK Python's default tracing exported OAuth access tokens, refresh tokens and client secrets into span attributes until a November 13 fix. A trace placed to see everything also stored everything.

What this means for builders

Three practices follow. First, list the paths an agent can act through, such as tool calls, shell, file system, network and child processes, and mark which ones each control actually sees. A gateway on tool calls says nothing about a script the agent runs. Second, check effects where they land: compare the database writes, files and requests that happened against those expected, as the rule-based evaluator did, instead of asking a model whether the transcript looks fine. Third, treat logs and traces as sensitive outputs with their own controls, because a record that captures every argument captures every secret.

This continues yesterday's note on moving permissions from pattern rules to capabilities. That was about how authority is granted. This one is about where checking happens and what each check can and cannot see.

What remains uncertain. guardscope's abstract does not report coverage numbers. PolicySniff needs a co-resident attacker on shared hardware and a particular runtime. The evaluator study's forbidden writes are a small sample of 11. None of these sources measure how often such gaps are exploited in production.

Analysis: At Belgavi's Book Store, Geetha could draw the refund agent's paths on one page: the tool API, the shell it uses for exports, and the outbound email connector. Each gets a mark for which control sees it. The nightly check should then compare the refund records actually written against the approved CASE list, rather than having a model read the day's transcripts.

Sources and evidence

Where a control sits decides what it can see. guardscope measures which of 15 agent action paths a runtime control checks, PolicySniff infers an agent's approval setting from shared-GPU timing, a rule-based check caught all 11 forbidden writes that LLM judges largely passed, and ADK tracing exported OAuth tokens by default. Map action paths, check effects where they land and protect traces.

Views are my own and do not represent any employer, past or present.