-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA256 NEFARIOUSPLAN-CANONICAL-V1 {"body_md":"## Cisco found the bug internally and published no endpoint\n\nCisco's advisory describes CVE-2026-20223 in one sentence:\n\n> A vulnerability in the access validation of internal REST APIs of Cisco Secure Workload could allow an unauthenticated, remote attacker to access site resources with the privileges of the Site Admin role.\n\nCWE-306, missing authentication for critical function. CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H, a 10.0. The \"Details\" section names no specific endpoint, no specific HTTP method, no parameter names, no schema. The discovery did not come with a researcher publishing alongside; Cisco found the bug during internal security testing. Cisco's fixed-version table names the patched releases; the patch diff is inside Cisco's source tree, not on GitHub. There is no public proof-of-concept.\n\nThe Hacker News writeup the PoC's README cites reproduces the same advisory text in different sentences. No additional technical detail. Cisco explicitly states there is no evidence of in-the-wild exploitation. The public record on 22 May 2026 contains the words \"REST API,\" the words \"Site Admin,\" and the CVSS vector.\n\n## The PoC was committed two days later\n\n`HORKimhab/CVE-2026-20223` shows two commits, both dated 22 May 2026: an initial scaffold and a `chore: CVE-2026-20223` one-liner. The repository contains a Python script, a Bash script, a README that wraps roughly thirty disclaimers around the words \"educational\" and \"academic,\" and a license. The Python script's class is named `CiscoSecureWorkloadPoC`. Its top-level workflow probes a list of guessed endpoints:\n\n```python\ndef test_unauth_site_admin_access(self):\n endpoints = [\n \"/api/v1/users\",\n \"/api/v1/roles\",\n \"/api/v1/sites\",\n \"/api/v1/admin/users\",\n \"/api/v1/internal/agents\",\n \"/api/v1/scopes\",\n \"/api/v1/policies\",\n \"/api/v1/system/info\",\n ]\n ...\n for endpoint in endpoints:\n resp = self.session.get(url, timeout=12)\n if status in [200, 201, 204]:\n self.logger.warning(\n f\"✅ VULNERABLE: {endpoint} returned {status} without authentication!\"\n )\n```\n\nThe Bash script adds `/api/v1/config` to the list and applies the same threshold. There is no version probe, no banner check, no response-schema validation, no JSON content sniff. The detection rule is one line:\n\n```python\nif status in [200, 201, 204]:\n```\n\nAn endpoint is reported `VULNERABLE` if it returns 200, 201, or 204 to an unauthenticated GET. A Cisco Secure Workload deployment's web tier has many endpoints that return 200 by design: public assets, JavaScript bundles, language packs, redirect targets, health checks, login pages, anything not gated behind the API authentication layer. The script does not check whether the response was authenticated. It does not check whether the response was meant to be authenticated. The condition for the `VULNERABLE` print is that the server answered.\n\nIf the threshold fires on a path, the script then attempts a POST against the same path with a hardcoded payload:\n\n```python\ndef create_admin_user(self):\n url = urljoin(self.target, \"/api/v1/users\")\n payload = {\n \"username\": \"cve202620223_admin\",\n \"password\": \"Poc123!@#\",\n \"role\": \"Site Admin\",\n \"email\": \"admin@compromised.local\",\n \"firstName\": \"CVE\",\n \"lastName\": \"2026-20223\"\n }\n resp = self.session.post(url, json=payload, timeout=10)\n if resp.status_code in [200, 201]:\n self.logger.critical(\"✅ Successfully created Site Admin user!\")\n```\n\nThe endpoint path, the JSON schema, the field names, the role-string format, and the success criterion are all guesses. The author has no patch diff, no traffic capture, no internal documentation, no working primitive. They have the advisory's sentence. The script reads that sentence and POSTs a payload that would, if the advisory's plain-English description happened to map exactly to a single-step user-creation primitive at `/api/v1/users` with a `role` key holding the literal string `Site Admin`, perform what the advisory describes. The advisory does not say any of those things.\n\n## The Site Admin string is the script's only specificity\n\nThree classes of artifact in the script attach it to CVE-2026-20223 rather than to any other unauthenticated-REST-API bug.\n\nThe class name `CiscoSecureWorkloadPoC`, the log prefix `CVE-2026-20223`, and the README title carry the CVE identifier. They name the artifact's subject. They do not constrain its behavior.\n\nThe endpoint list reads as Cisco-Secure-Workload-shaped. `/api/v1/sites`, `/api/v1/scopes`, `/api/v1/policies`, and `/api/v1/system/info` look like paths a multitenant policy product would expose. There is no public Secure Workload OpenAPI document the author could be matching against; the names match the marketing-page vocabulary of zones, scopes, and policies that Cisco uses to describe the product. A defender with access to a real Secure Workload deployment can confirm by comparing the script's path list to the product's actual admin API guide. The script's relationship to the real REST surface is the relationship between the words on a product page and the routes a backend serves.\n\nThe POST payload's `\"role\": \"Site Admin\"` is the entire claim to specificity. The string is correct in one sense: Cisco Secure Workload's RBAC schema does have a Site Admin role, the advisory spells it that way, and the product does. The string is also where the author's information ends. They do not know whether the role is stored under a `role` key, a `roles` array, a `role_id` integer, an `assignedRoles` collection, a `permissions` set, or a product-specific data model entirely. They do not know whether user creation is a single POST or a multi-step flow that requires a tenant ID, a scope reference, or a session token first. They do not know whether the role-name is passed by display string or by an internal slug.\n\nWhat they have is the string from the advisory. They typed it into the payload. The payload's job is not to create a Site Admin user. The payload's job is to be the payload of a script with the CVE number in its filename.\n\n## Post-advisory sympathetic magic\n\nThis blog has named the [Placeholder PoC](/patterns/placeholder-poc) pattern five times. The exhibits cover repositories where the author's lack of the underlying primitive is admitted in a comment ([InSAT MasterSCADA](/posts/insat-masterscada-cve-2026-22553-placeholder-poc) marked the field `# placeholder field`), where the script's threshold matches the product rather than the bug ([the curtishoughton CVE-2025-20362 scanner](/posts/cisco-asa-cve-2025-20362-scanner-detects-the-logon-page) fires `HIGH RISK` on every Cisco WebVPN that serves its logon page), where two byte-identical scripts stop at the function the real exploit asserts is not yet root ([the Fragnesia placeholders](/posts/cve-2026-46300-fragnesia-placeholders-stop-at-verify-write-denied)), where the script's surrounding documentation describes a project that does not exist ([the kaleth4 CredSSP repository](/posts/credssp-cve-2025-47987-poc-documents-a-different-project)), and where the artifact is a slide deck the students never claimed was an exploit ([the Tomcat coursework](/posts/tomcat-cve-2025-24813-the-poc-is-a-slide-deck)).\n\nCVE-2026-20223 adds a sub-shape under that pattern: the post-advisory sympathetic-magic scanner. The structural conditions that produce it: a vendor found a bug internally, published a CVSS 10.0 advisory describing the impact in plain English without specifying the primitive, and the public record contains the advisory and the CVE identifier. No third-party researcher has the patch diff. There is no other PoC to crib from. The PoC author has the advisory paragraph as their entire source of truth. They construct a script whose every concrete claim is generated by reading the advisory and guessing what it implies: the endpoint list, the JSON schema, the role-string format, the success criterion.\n\nThe sub-shape is distinct from the curtishoughton ASA case along two axes. The ASA scanner had two hundred and thirty-three days of public discourse to mine: KEV listing, multiple Cisco advisory updates, an in-wild attack-variant disclosure, joint government attribution to a named threat actor cluster. None of that work appears inside the script, but the surrounding ecosystem produced a body of public material the author could in principle have drawn on. This case has 48 hours of public existence. There is no surrounding body of material. The advisory is the entire source.\n\nThe ASA scanner's specificity is its keyword list: `webvpn`, `cscoe`, `anyconnect`, words Cisco's portal HTML contains by design. This script's specificity is the string `Site Admin`. Both are facts pulled out of vendor-authored text and treated as the script's claim to relevance. The mechanism by which the script attaches itself to its CVE is, in both cases, sympathetic: a textual correspondence between strings in the advisory and strings in the script substitutes for the absent primitive.\n\n## What a defender running the scanner sees\n\nA defender running `HORKimhab/CVE-2026-20223` against a patched Secure Workload deployment sees `VULNERABLE` outputs on every path in the list whose handler returns 200, 201, or 204. Cisco's REST surface is large; healthy deployments respond on many of these. The defender escalates. They hand the log to a vendor support engineer, who reads the script, sees that the detection threshold is \"the server answered,\" and closes the ticket. The defender's report becomes noise in the support queue.\n\nA second defender, evaluating CVE-2026-20223 exposure on an actually-unpatched deployment, sees the same `VULNERABLE` output on the same paths. They report the exposure as confirmed. The actual vulnerable endpoint, whichever one it is inside Cisco's REST surface, is not in the script's path list. The defender's confirmed-exposure report describes a finding the script did not make.\n\nA scanner pipeline polling GitHub for CVE-numbered repositories indexes `HORKimhab/CVE-2026-20223` against CVE-2026-20223 and registers a hit. The defender community downstream of that pipeline now sees CVE-2026-20223 as having a public PoC. The actual public exploit, if and when it appears, ships into a noise floor where this script already exists as the canonical \"PoC.\"\n\nCisco found the bug. Cisco did not publish the endpoint. The PoC author did not have the endpoint either. The Cisco advisory has two distinguishing facts. The script has those two facts and a loop around them.\n\nPoC: [HORKimhab/CVE-2026-20223](https://github.com/HORKimhab/CVE-2026-20223)","closing_line":"The advisory was the input. The advisory was also the output.","hook_md":"Cisco published advisory `cisco-sa-csw-pnbsa-g8WEnuy` for CVE-2026-20223 on 20 May 2026: a CVSS 10.0 missing-authentication fault in Cisco Secure Workload's REST API, allowing an unauthenticated remote caller to access site resources with the privileges of the Site Admin role. The bug was found during internal Cisco security testing; no patch diff, no exploit detail, no endpoint name. Two days later, `HORKimhab/CVE-2026-20223` was committed to GitHub. The script inside it probes nine guessed `/api/v1/*` paths and POSTs a payload whose role field is the literal string `Site Admin`. The two distinguishing facts in the Cisco advisory are also the only specifics in the PoC.","post_id":434,"slug":"cve-2026-20223-the-advisory-was-the-poc","title":"CVE-2026-20223: The Cisco Secure Workload PoC Is The Advisory With A Loop Around It","type":"initial","unreadable_sentence":"The payload's job is not to create a Site Admin user. The payload's job is to be the payload of a script with the CVE number in its filename."} -----BEGIN PGP SIGNATURE----- iHUEARYIAB0WIQRf0htP5+SjynlxywneZjl4jgkQJgUCamTy+wAKCRDeZjl4jgkQ Jhz2AP0YSADjzl+ldurNTZ6YPjlbZv2XnIzojcH43WpfuHFMKQD+LV1UlSG0BSxI HuvMnkOY1RLIHI+N5whmZ9ZfpDyBnw0= =/K5B -----END PGP SIGNATURE-----