# Agent investigation is only as checkable as the evidence record it leaves behind

Guide, 19 Nov 2026, read 19 Nov 2026

**The short answer.** A checkable agent record names the finding, the access the agent held, each source with a timestamp, each test with its result, the conclusion and every unknown. A complete sample for CVE-2026-82329 in Artifactory follows, with a reviewer's replay steps and four vendors' own statements about showing evidence.

## A record has seven parts, and each part fails in its own way

A conclusion alone is a claim. A record someone else can replay holds the parts below. The agent or the platform writes parts one to six, and a named person writes the seventh.

The four-question tool on the home page sends a finding to agent investigation when a control sits in front of the service, or when two or three of code called, service reached and control in front are not sure. Those are the findings whose answer sits in configuration, logs and files, which is what this record holds.

| Part | Written by | Usual failure |
| --- | --- | --- |
| 1. Finding and asset | Platform, from the scanner | Asset identity is a name that was reused |
| 2. Access held | Platform, from the grant | Lists the role requested, not the access in force |
| 3. Sources read, with time | Agent | Gives no time, so a stale read looks current |
| 4. Tests against advisory preconditions | Agent | Cites no advisory text, or tests a different condition |
| 5. Conclusion and scope | Agent | Says 'not exploitable' with no asset and date |
| 6. Unknowns, with the reason | Agent | Omitted, so a gap reads as a pass |
| 7. Review and expiry | Human reviewer | Missing, so the closure never expires |

## A complete sample record for CVE-2026-82329 on one Artifactory instance

The question here is not only 'is the version affected' but 'was this instance compromised', which is why CVE-2026-82329 suits agent investigation. Fastly reports disclosure on 28 Aug 2026, about 171,000 attempts on 1 Sep 2026 and a peak near 406,000 on 2 Sep 2026. The instance, hosts, dates and counts in the record are an illustration. The advisory facts are real.

Parts one and two come first.

| Field | Entry |
| --- | --- |
| Record | INV-4417, version 1, opened 19 Nov 2026 09:12 UTC |
| Finding | CVE-2026-82329, JFrog Artifactory Access service join-key authentication bypass, CVSS 9.8, added to KEV on 2 Sep 2026, first mass exploitation 1 Sep 2026 (Fastly) |
| Asset | artifactory-prod-1, self-hosted, owner Platform Engineering |
| Access held | Read-only API token with scope for system version and user listing; read-only search on the reverse-proxy log index. It holds no shell, file-system or write access |
| Access granted by | Platform Engineering lead, 3 Nov 2026, expires 3 Feb 2027 |

## The sample record's tests each name a precondition, a source and a time

The advisory preconditions are a self-hosted release in an affected range, a default install with no additionalJoinKeys set, and network access to the join endpoint. On a default install the Access service trusts an empty join key. The attack, as Fastly describes it, is a request to POST /access/api/v1/registry/join carrying a forged token whose key ID is the SHA-256 of the empty string, e3b0c442...b855. Fastly calls an HTTP 201 tied to that key ID "high-confidence evidence of compromise".

| Test | Source, time (UTC) | Observed | Result |
| --- | --- | --- | --- |
| T1 Release affected now | GET system version, 19 Nov 2026 09:14 | 7.146.38 | Fixed build. JFrog's table lists 7.146.0 to 7.146.36 as affected and NVD treats every release below 7.146.38 as affected. 7.146.37 is in neither JFrog's affected list nor its patched list, so only 7.146.38 counts as fixed |
| T2 Release affected during exposure | Change ticket CHG-9120, read 09:15 | Upgrade from 7.146.30 to 7.146.38 closed 9 Sep 2026 | Affected from 28 Aug to 9 Sep 2026 (12 days); 7.146.30 is inside both JFrog's and NVD's range |
| T3 Join endpoint reachable | Reverse-proxy routing config as exposed in the log index, 09:17 | Route to /access/ open to the corporate VPN range only | Not internet reachable; reachable from the VPN range |
| T4 Join requests with an HTTP 201 | Log index search, join path, status 201, 09:21 | Retention starts 21 Sep 2026. 0 matching requests from 21 Sep to 19 Nov 2026 | No evidence of compromise in the retained window |
| T5 Unexpected admin users | User listing, 09:24 | 4 admins, all on the owner's roster | No unexpected admin |
| T6 Groovy user plugins | Needs file-system access | Not run | Unknown |
| T7 Join key rotated and tokens revoked since 28 Aug 2026 | Change tickets, search 'join key', 09:27 | No ticket found | Not done on record |

## The sample record's conclusion states its scope and lists its unknowns

- Conclusion: artifactory-prod-1 is not vulnerable on its current release (T1). It was on an affected release from 28 Aug to 9 Sep 2026 (T2) and was reachable only from the VPN range (T3), so exposure to internet scanning was limited.
- Compromise: unknown for 28 Aug to 20 Sep 2026, because the log index starts on 21 Sep 2026 (T4). No sign of compromise from 21 Sep 2026 onward (T4, T5).
- Unknown, plugins: the agent holds no file-system access, so T6 was not run.
- Open action for the reviewer: the ledger mitigation says patching does not revoke tokens already issued and advises rotating the platform join key and revoking tokens issued since 28 Aug 2026. T7 found no record that either was done.
- Status: needs human action, not closed. Reviewer: Platform Engineering lead. Expiry of any later closure: 19 Feb 2027, or on a new advisory revision.
- Version history: v1 only. A later version adds a row below with its date and the evidence that changed.

## A reviewer replays a record in five steps

The same five steps work on any vendor's record. Run them on a record for a finding you already understand, and include one you expect to come back unknown.

- Open the advisory and confirm the record quotes a precondition from it, not a paraphrase of the title.
- For each test, repeat the source read yourself, or confirm the source and time are specific enough that you could.
- Check the access list against the sources: every source read must fall inside the access held. T6 above shows a source outside it, correctly marked unknown.
- Look for the unknowns. A record with no unknown on a finding whose evidence needs logs, files or hosts the agent cannot read has hidden a gap.
- Confirm a person is named for the closure and that an expiry or trigger is set. The agent reads, and people decide on closures that become exceptions.

## Four vendors describe how their products show evidence

These are the vendors' own statements, in alphabetical order. They describe design, not measured accuracy, and none comes with a sample record.

- Cogent: "Every output includes the full chain of logic: data sources queried, agents involved, and why this path was selected over alternatives." When signal is short it says it "will surface unknowns/low confidence".
- Maze: "Every recommendation comes with the evidence and reasoning behind it."
- Onit: a Reachability Agent marks each exposure Reachable, Not Reachable or Unknown, and "when evidence is missing, an agent returns 'unknown' with the reason".
- Seemplicity: AI Analysts investigate each asset by analyzing "live runtime configurations, exploit prerequisites, network reachability, code and dependency usage, and compensating controls".

## Sampling closures measures your own deployment

Pull ten findings closed as not exploitable each month, read each record with the five steps above and count how many the reviewer reopens. That count is specific to your deployment. Keep the agent's access list under change control too, since a record that cites a source the agent can no longer reach cannot be replayed.

## Related pages

- [Exploitable vs vulnerable](/guides/exploitable-vs-vulnerable/)
- [Triaging an internet-facing appliance finding](/guides/triaging-an-internet-facing-appliance-finding/)
- [How to prove a finding is not exploitable](/guides/prove-a-finding-is-not-exploitable/)

## Sources

- Fastly on CVE-2026-82329, 3 Sep 2026: https://www.fastly.com/blog/cve-2026-82329-jfrog-artifactory-authentication-bypass-exploitation-activity (read 19 Nov 2026)
- Cogent AI engine page: https://www.cogent.com/ai/cogent-ai-engine (read 19 Nov 2026)
- Maze cloud platform page: https://mazehq.com/platform/cloud (read 19 Nov 2026)
- Onit blog on AI agents in vulnerability management: https://onit.security/blog/ai-agents-vulnerability-management/ (read 19 Nov 2026)
- Seemplicity AI Analysts announcement: https://seemplicity.ai/news/ai-analysts-autonomous-vulnerability-response/ (read 19 Nov 2026)
- JFrog security advisories (CVE-2026-82329 affected and patched versions): https://docs.jfrog.com/releases/docs/jfrog-security-advisories (read 19 Nov 2026)
- NVD record for CVE-2026-82329: https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-82329 (read 19 Nov 2026)
- CISA KEV entry for CVE-2026-82329 (added 2 Sep 2026): https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-82329 (read 19 Nov 2026)

Last reviewed 19 Nov 2026. Canonical: https://reachabilityanalysis.com/guides/how-agent-investigation-records-its-evidence/
