Skip to main content
A finding is one place where what an agent declared it would do and what the machine saw it do came apart. The Findings page lists them; this is the path from a row to a decision. Setting a verdict needs the Policy Writer role or higher.
1

Pick the finding

Open Findings. Filter by severity (critical, high, medium, low), by family (intent vs truth, supply chain) and by verdict (Unjudged is where new work is). In the Findings panel choose Flat and sort By severity to work from the top.Each row shows Severity, Finding (name and type id), What happened, Score, Verdict, Session and Age.
2

Read what was declared and what was observed

Open the row. The Overview tab pairs the declared action with the observed one; the Evidence tab shows Detector evidence: the recorded facts the finding rests on. Raw shows the same as data.Framework mappings names the controls this type of finding maps to. The page says it plainly: “it is not a compliance opinion.”
3

Follow it into the session

Use Open session or View in timeline. The session opens on Investigate: Activity lists its recorded actions, Conversation what was said. Context holds Session info (working directory, agent, platform, model, parent session) and Child sessions; Behavior holds Process lineage when the sensor tied processes to the session. Look at what happened just before and after the finding.
4

Check the machine

Open the session’s Host link. The host page shows Last heartbeat in its header, the capture lanes and install checks the sensor reported on Capture, and its agents on Agents. A degraded host, or a Capture tab saying “Lane states are last-known readings”, may be missing evidence, which changes how much weight the finding carries.
5

Ask Sentinel if the picture is unclear

Investigate in Sentinel asks Sentinel about the finding in a new conversation beside the page. Sentinel answers from your workspace’s own rows and shows every query it ran. See Ask Sentinel.
6

Record a verdict

Back on the finding, set True positive, False positive or Suspicious. Verdicts are filters on the list, so an unjudged queue stays short. Export CSV exports the current list.
7

Track it, or draft a rule

Only a critical alert opens an Incident on its own. For anything else you want tracked to resolution, open Incidents and work the record. The panel beside it has three menus: Status (Open, Investigating, Resolved; an older record may show Contained or Closed), Owner and Verdict. The button at the top offers the next step: Assign to me while nobody owns the record, then Set verdict. The ⋯ menu holds Copy link, Copy id, Open the first session and Status…. Add a note under History with Add note. Assigning to another person needs Admin or Owner.The panel also links Draft a shadow rule, which opens Policies › Rules, where you draft the rule. See Write and publish a rule.

What the risk names mean

Every page uses these names. On the Overview and in Top detectors on the Findings page, types outside this list are counted as “Other findings”; a single finding of such a type shows its type id. On macOS, in an upcoming agent release, Command differs from declared judges the program that ran (under a declared ls, a shell starting curl is reported; a test runner starting git is not), and Undeclared command names that program. With that release, when an agent runs a Python tool by its own name, such as pytest, pip install or uv run pytest, the further copies of Python it starts (test workers, package builds) are not recorded as a match, as is already the case for python3 -m pytest. On Homebrew and Apple framework Python this removes matches earlier versions recorded; anything of another kind those copies start, such as curl, is still recorded. A versioned interpreter such as python3.12 that a build tool starts as a separate process, for example cd pkg && python3.12 -m pytest in a make test recipe, is newly recorded when the command does not name it; a recipe line the build tool runs in place is not, as before. There is at most one match per program per command, and at most 32 programs per command. Which programs a command counts as naming, including the limits for inline commit messages, is described in How a command is read.

Tool supply-chain findings

First-time checks, invisible characters and cross-references arrive in an upcoming agent release. Not checked: tools in OpenAI Chat, Bedrock Converse or Gemini formats; text inside a parameter schema.

If the list is empty

“No divergences in this window.” is a result for the window you chose. A note that counts are unavailable is not: it is a coverage gap, not a quiet fleet. Widen the window, then check Coverage on the Overview and the Silent filter on Data Sources before concluding the fleet was quiet.

Next

Last modified on November 15, 2026