What shipped, and what it is evidence of
On November 5, 2026 Anthropic added Security posture data to the Compliance API, in beta for Claude Enterprise organizations. Two pieces arrived together: the Activity Feed now records changes to attestations on the Security posture page in claude.ai Organization settings, and a new endpoint, GET /v1/compliance/organizations/{organization_id}/security_posture, returns the organization's current results 1.
The page itself compares your organization's Claude settings with Anthropic's recommended security controls. It reports on your settings and does not change them 2. That framing matters for audit. This is configuration evidence: is single sign-on on, which controls did someone confirm by hand, which risks were accepted. It says nothing about which applications called which model, with which key, carrying which data. Anthropic is explicit that the security center is guidance based on its recommendations, not a certification or a guarantee of compliance with any law or standard 3. Do not present it to an auditor as either.
The rest of this guide covers what the data contains, what an auditor can and cannot conclude from it, and a repeatable export for an access review.
What the data contains
Security posture evaluates a list of controls and gives each one a result. A control is one recommended setting or practice, with an identifier that is the same for every organization, a display name, and sometimes a current value and a recommended value. There are three kinds 2:
- Automatic controls: an automatic check of the current value against the recommended value, for example
sso_enabled. - Manual controls: decided by a person, because Anthropic cannot read the setting (for example when it lives in another vendor's console), such as
oauth_conditional_access. - Information-only controls: no recommended value, so they are reported and not rated, such as
allowed_invite_domains.
Each control ends up with one of four results: passed, failed, not_attested (a manual control with no attestation in effect) or descriptive (not rated) 2.
On top of results sit attestations: a person's recorded statement of how the organization stands on a control. The values are completed (in place), accepted_risk (not in place, risk accepted) and control_failed (not in place). The API reads attestations and cannot make them 2.
Two details change how you read the numbers:
- An accepted risk in effect counts as Passed on an automatic or manual control, though the control is not in place 2. A green result can therefore be an exception someone signed.
- An attestation ends when someone clears or changes it, when its end date (
review_date, which the attester sets to 90 days or none) arrives, or, on some controls, when the current value changes or the check passes 2.
The endpoint returns each control's id, display_name and a result in rating. It reads one organization per call and returns no history 2. It also returns no comments: attestation comments, up to 2,000 characters, are stored with the attestation, but neither the endpoint nor the Activity Feed exposes them 2.
Who can read it, and where it returns nothing
Availability is narrower than the Compliance API as a whole. Security posture needs a Claude Enterprise organization whose parent organization has the Compliance API enabled, plus a Compliance Access Key with the matching scope. For other plans and for Claude Console organizations the Activity Feed records no Security posture activities and the endpoint returns 404. It also returns 404 if the page was turned off for the organization, even for a key at the parent level, and for some organizations even when every condition holds 2.
The scopes are read:compliance_org_data or read:org_audit for the endpoint, and read:compliance_activities or read:org_audit for the feed's Security posture activities 2. Authentication uses an x-api-key header and an anthropic-version header on https://api.anthropic.com; all /v1/compliance/* endpoints share a rate limit of 600 requests per minute per parent organization 4. A daily snapshot of a few organizations is nowhere near that.
If your Claude usage is mostly Console API traffic, treat a 404 as a finding to record, not an error to retry: it means this evidence source does not exist for that organization.
What an auditor can use it for
Map each audit question to what the data actually answers.
| Audit question | Answered by Security posture? | Where the evidence comes from |
|---|---|---|
| Is SSO enforced for the Claude Enterprise organization? | Yes, if sso_enabled is an automatic control in your results | security_posture endpoint, dated snapshot |
| Which controls rely on a person's confirmation? | Yes: manual controls and attestations | Endpoint plus Activity Feed |
| Who accepted which risk, when, and until when? | Yes | Activity Feed attestation activities; the endpoint for risks in effect |
| What was the posture on a past date? | Only from your own stored snapshots and feed | Your archive; the endpoint has no history |
| Which employees can sign in and with which roles? | Partly, by other Compliance API endpoints | Organizations, users, roles and groups endpoints |
| Which application called Claude, with which key, with which data? | No | Gateway or application logs |
| Was a prompt blocked before it reached a model? | No | Inference hooks logs or gateway guardrail logs |
The Activity Feed gives the change history: it records an activity when someone attests to a control, changes an attestation or clears one, and sometimes when one ends automatically 2. Activities are queryable within 1 minute of occurring, except that Security posture activities can take longer, and they are retained for 6 years. Recording is not retroactive: it starts when the Compliance API is first enabled, with no backfill 5. If the Compliance API was enabled last week, an auditor asking about last quarter gets nothing from this source.
Anthropic also lets owners download the security center page as a report for auditors; it marks which controls were checked automatically and which an Owner confirmed 3. Use that report as the human-readable cover sheet and the API snapshot as the machine-readable record.
NIST's AI RMF is a voluntary framework, and NIST published a Generative AI Profile (NIST-AI-600-1) on July 26, 2024; the page also notes that AI RMF 1.0 is being revised 6. If your program maps evidence to it, this data supports the "how is the tool configured" side of a mapping, not the monitoring of what the tool does.
Gaps that need gateway-side logs
The posture data stops at the claude.ai Enterprise settings boundary. These questions need evidence from somewhere else.
- API traffic by caller. Anthropic's Usage and Cost Analytics API (beta) reports token usage and cost for Console and Enterprise organizations, grouped by dimensions such as model, workspace, API key and the user or service account 1. That is aggregate usage, not a per-request record with your application's own context.
- Per-request decisions. Inference hooks, now out of beta for Enterprise, act inline: your security server receives each governed prompt before inference and can deny it. The Compliance API, by contrast, retrieves records after the fact 4. Hooks cover Claude products in the Enterprise organization; they are not a log of what your own backend sent through an API key.
- Content retention policy. The Compliance API can return chats, files, projects and session transcripts for Enterprise 4, which is the opposite problem: it can hold content. Decide whether the audit pack should include any of it, and who may see it.
- Keys and spend ceilings you set outside Anthropic. Which gateway key belongs to which project, who may use it, what ceiling applies and when it was rotated are facts only your gateway or secrets manager knows.
For API traffic that goes through a gateway you run, the gateway's request log is the record that answers "who, which model, which key, how much, when". On INFRO, the shipped pieces that fit this are the audit log of member, key and policy changes with actor and timestamp, the request log and usage series exports as CSV, per-request traces with payload logging off by default and a retention window you choose, and project-scoped access for members. Their request and response contract is in the INFRO docs (opens in a new tab) 7. INFRO sees only traffic that goes through it, so it complements the Anthropic data and does not replace it. For the key-level side of the review, see the AI access review checklist.
Export it into an access review
The aim is a dated, reproducible evidence bundle. Run this on a schedule (daily or weekly), not on the day the auditor asks.
- Provision a read-only key. Create a Compliance Access Key in claude.ai with only
read:compliance_org_dataandread:compliance_activities, or the singleread:org_auditscope if you accept its breadth. Keep it in your secrets manager, not in a notebook. - Snapshot current results. Call the endpoint for each organization and store the raw response with a timestamp. Because there is no history, the stored snapshots are the history.
- Pull the feed incrementally. Page the Activity Feed with a stored cursor, filtering to the Security posture activity types listed in the API reference, and persist the cursor only after the page is stored 5.
- Join, then diff. For each control, join the latest result with its latest attestation and flag three cases:
accepted_riskin effect, an attestation whosereview_datehas passed, and a result that changed since the last snapshot. - Sign the review. The reviewer records approve, remediate or accept for each flagged control. That decision is your evidence; the snapshot is only the input.
A minimal snapshot script follows. The response key holding the control list is an assumption here: confirm the exact shape against Anthropic's results reference before relying on it.
import json, os, datetime, requests
ORG_ID = os.environ["CLAUDE_ORG_ID"]
URL = f"https://api.anthropic.com/v1/compliance/organizations/{ORG_ID}/security_posture"
HEADERS = {
"x-api-key": os.environ["ANTHROPIC_COMPLIANCE_ACCESS_KEY"],
"anthropic-version": "2026-10-21",
}
resp = requests.get(URL, headers=HEADERS, timeout=30)
if resp.status_code == 404:
# Not Enterprise, page turned off, or not available to this org: record it.
record = {"org": ORG_ID, "status": "unavailable_404"}
else:
resp.raise_for_status()
body = resp.json()
# ASSUMPTION: confirm the list key against the results reference.
controls = body.get("controls", body.get("data", []))
record = {
"org": ORG_ID,
"status": "ok",
"counts": {},
"raw": body,
}
for c in controls:
r = c.get("rating")
record["counts"][r] = record["counts"].get(r, 0) + 1
record["taken_at"] = datetime.datetime.now(datetime.timezone.utc).isoformat()
stamp = record["taken_at"][:10]
with open(f"posture-{ORG_ID}-{stamp}.json", "w") as f:
json.dump(record, f, indent=2)Store the files somewhere write-once, so the snapshot is credible as a record of that day.
Pitfalls to check before you show it to an auditor
- The page and the endpoint can disagree. The response leaves out some controls and some attestations, so the page, the endpoint and the feed do not always show the same thing. Anthropic documents how an attestation shows in the response and how to check stored events against current results 2. Run that check once, and note any controls missing from the endpoint in the evidence pack.
- Beta means the shape can move. The feature is in beta 2. Pin the
anthropic-versionheader, store raw responses, and keep your join logic separate from the fetch step so a field change breaks one function. - A green result is not a control test. Because accepted risk counts as Passed, report passed, failed, not attested and accepted-risk counts separately. A summary that merges them overstates the position.
- Feed lag. Security posture activities can take longer than a minute to become queryable 5, so do not alert on the absence of an activity for the first few minutes.
- Scope of claim. State in the pack that the evidence covers Claude Enterprise organization settings, name the organizations included, and list the ones that returned 404.
The result is a short, honest pack: a dated posture snapshot, the attestation history with approvers and end dates, a list of accepted risks, and a pointer to your gateway or application logs for request-level questions.
