Reading a report

Every activity produces a Markdown report in a fixed shape. Knowing the shape lets you read one in a minute and judge how far to trust it.

Where to find a report

WhereHow
DashboardOpen an assessment, then a run. The report is rendered with severity badges and, after secfoo vendor mermaid, diagrams. The run page also lets you download report.md and prompt.md.
Terminalsecfoo show <run-id>. Use secfoo list to find the run ID.
Disk~/.secfoo/reports/<run-id>/report.md, next to the prompt and the agent's logs.

Start at the top: summary and risk rating

Every report opens with a short executive summary and a line like:

**Overall risk rating:** High

That's the agent's judgement of the target as a whole. It isn't simply the worst finding: one high finding in a test helper may still leave the overall rating at Medium. For Responsible AI Compliance, this line becomes the assessment's risk badge.

The register: every finding at a glance

Right after the summary comes a table of every finding, most severe first. Activities add their own columns, but each row has an ID (F1, A3, T7, D2, S1…), a short title and a severity. Use it to decide what to read in full.

SeverityMeaning
CriticalExploitable now with serious impact. Fix before anything else.
HighSerious, and likely exploitable. Fix soon; good candidates for a CI gate.
MediumReal weakness with limited impact, or needing conditions to exploit.
LowHardening and hygiene.

Each finding in detail

Below the register, each finding gets its own section headed ### [SEVERITY] ID: title, with:

LocationFile and line, component, or N/A for a design-level finding.
DescriptionWhat the issue is and why it matters.
RecommendationA concrete fix, not "review this". SAST reports add an illustrative diff.
CWEA specific CWE only when the agent is confident. unverified means it wasn't.
VerdictConfirmed: exploitable as-is. Conditional: exploitable given a stated precondition. Latent: a real defect that isn't currently reachable.

The severity counts secfoo prints after a run, and that --fail-on gates on, come from these headings.

Activity-specific verdicts

ActivityLook forValues
Security Architecture ReviewDesign verdict, in the last sectionSound · Sound with conditions (conditions listed) · Not sound
Threat ModelingDisposition column in the Threat RegisterMitigated · Gap · Accepted · Transferred. A Mitigated threat with test reference none is a claimed but unproven mitigation.
SCAReachability column in the Risk RegisterReachable · Conditionally reachable · Not reachable
Secret ScanningValidity column in the Findings RegisterLooks live (rotate now) · Unclear · Placeholder
Third-Party RiskRecommendation sectionApprove · Approve with conditions · Needs more documentation · Reject

Read the coverage notes

The last section says what the agent reviewed, what it skipped or couldn't reach (a Confluence page without an MCP connection, an excluded folder, files it ran out of time for) and how confident it is. A clean report with thin coverage notes means less than one that says it read everything. At quick depth, expect the agent to say it sampled.

How far to trust it

A report is a capable reviewer's first draft, not a verdict. The briefs push agents to be honest: to write unverified instead of guessing a CWE, never to invent a CVE ID, to give each finding a location and a concrete fix, and to say what they didn't cover. secfoo computes CVSS scores itself and checks dependencies against OSV.dev rather than trusting the model. Even so:

  • Check each finding against the code before acting on it, especially Conditional ones.
  • Record your decision. In the dashboard you can mark a SAST finding confirmed, a false positive, or an accepted risk. See Recording risk decisions.
  • Expect some variation between runs. The report's shape is fixed, but the wording isn't, and a later run may surface a finding an earlier one missed. SAST, SCA and secret findings are matched by the code they point at, not their wording, so rephrasing doesn't create duplicates.
  • See exactly what the agent was asked. Each run's prompt is saved as prompt.md, downloadable from the run page.
← Activity reference Using the dashboard →