On November 18, 2026, CISA added CVE-2026-90817 for Vanderbilt REDCap after in-the-wild exploitation of an unauthenticated remote code execution path in survey passthrough and Data Import logic. Affected builds run REDCap 13.3.0 and higher until 16.0.49 LTS, 17.3.10 LTS, or 17.4.4 Standard. Attackers need a valid public survey hash, not a user login. Securifera’s walkthrough explains how decode mismatches, NOAUTH context, and PHP stream filters chain into Blade template cache code execution.
Research platforms that also run coding agents for analysis scripts should patch REDCap first—agents do not protect the PHP app boundary.
TL;DR
| Question | Answer |
|---|---|
| Exploited? | Yes — KEV 2026-11-18 |
| Auth? | No account; survey hash required |
| Patch | 16.0.49 / 17.3.10 / 17.4.4+ |
| Federal due date | 2026-11-21 |
| Forensics | Required per KEV row |
Who runs REDCap next to agent tooling
REDCap hosts CRFs, survey links, and data exports for clinical trials and social science. The same universities deploy Claude Code or Copilot on staff laptops that SSH to REDCap VMs for one-off exports. A REDCap RCE is therefore both a PHI incident and a developer workstation pivot if SSH keys live on the server.
November 2026 KEV should trigger cross-team war rooms: IR, research compliance, dev platform, and AI enablement must not treat this as only a vendor ticket.
Attack chain (from Securifera)
- Attacker obtains an active public survey link (hash in URL).
- Survey passthrough (
__passthru) decoding differences route the request to general project dispatch instead of the survey-only surface. NOAUTHcontext skips normal CSRF checks while granting data import privilege miscalculation.- Attacker POSTs to Data Import with crafted
fnameusing PHP filters (read/write primitive). - Write lands in predictable BladeOne cache path; include executes attacker code.
GitHub GHSA-hg88-qg2r-c8mx mirrors the same description NVD uses.
Patch, containment, and detection
Upgrade per vendor LTS tracks above.
If upgrade slips:
- Disable or restrict public surveys not strictly required (NVD-referenced guidance).
- Block survey-context access to DataImportController and ParcelController at WAF where possible.
Hunt signals (Securifera):
- Encoded
__passthruparameters DataImportController:indexfrom survey referrersfnamevalues withphp://,zip://, or filter chains- Changes under
temp/cache - Web-server child processes from PHP user
What this means if you run AI coding agents
Academic and hospital IT increasingly let agents draft REDCap exports, R scripts, or Python cleaners on shared app servers. Compromise of REDCap yields PHI and study data, not just the agent session.
| Risk | Mitigation |
|---|---|
| Agent stores API keys on REDCap host | Rotate after patch; use local secret stores |
| Agent curl tests hit survey URLs | Block agent-driven fuzzing against production surveys |
| Shared VM | Segment REDCap from arbitrary agent shell access |
Beam watches agent tool use on the workstation; it does not harden REDCap PHP. Pair compliance audit trails on agents with REDCap vendor patching.
Related reading
- Rails Active Storage ImageMagick CVE
- FastAPI generator import RCE
- November attacks timeline
- First 24 hours after a vendor breach
PHP filter chain and Blade cache sink (technical depth)
Securifera’s write-up is the authoritative mechanism article for defenders who cannot wait on a vendor PDF. The chain abuses three architectural layers that looked safe in isolation:
Survey passthrough decode mismatch: Public surveys use __passthru to reach internal routes. When URL decoding on the allowlist and the router disagree, a request that should stay inside survey-only handlers reaches the project dispatcher—the same routing surface authenticated project users hit.
Rights calculation in NOAUTH context: With NOAUTH defined for public surveys, CSRF checks and privilege assignment order change. The attacker receives data_import_tool capability without a REDCap login—Securifera’s table maps each defect to a fixable component.
Data Import fname trusts client paths: Attackers supply php://filter or similar stream wrappers so the import handler reads and writes attacker-chosen resources. BOM normalization turns a read into overwrite of predictable BladeOne compiled templates under temp/cache. Including that template executes PHP as the web user.
Survey hash precondition: CVSS still shows no privileges required because the hash is public-by-design on survey links. Securifera notes discoverability via search engines and archival scanners—treat any live public survey as credential-like exposure for this bug.
Version matrix
| Track | Patched release (per NVD/OpenCVE) |
|---|---|
| 16.x LTS | 16.0.49 |
| 17.3 LTS | 17.3.10 |
| 17.4 Standard | 17.4.4 |
Affected floor: 13.3.0+ below those builds. Identify version from REDCap Control Center or REDCAP_VERSION constants in redcap_connect if you operate multi-tenant installs.
KEV forensic triage (BOD 26-04)
Unlike LiteLLM’s KEV row, REDCap’s marks forensic triage required. For covered agencies—and as a private best practice—collect:
- Web logs 90 days back for
DataImportControllerhits from survey referrers temp/cacheintegrity snapshots- New admin accounts or cron entries on the host
- Outbound connections from PHP-FPM during survey hours
If you find indicators, assume study data and PHI exposure; invoke breach runbook rotation steps for database creds, OAuth, and S3 backup keys on the same server.
Institutional response playbook
University security offices often learn REDCap patches from central IT, not from PI laptops running agents. Assign owners:
- Application owner: schedule LTS upgrade window.
- IR team: run forensic triage if KEV applies to your sector.
- Research IT: pause new public surveys until patched.
- Agent users: no production fuzzing on live survey URLs.
Communicate to researchers that November 2026 KEV means assume breach if the host was internet-facing on vulnerable builds—rotate integration passwords to EDC vendors, LDAP bind accounts, and object storage credentials REDCap uses.
Coding agents on research infrastructure
| Scenario | Risk |
|---|---|
| Agent writes PHP export scripts on REDCap host | Shared compromise domain |
| Agent curl-fuzzes survey URLs during “security review” | Accidental weaponization evidence |
Agent schedules cron cleanup under temp/cache | Can mask attacker template drops |
Use read-only agent roles on production REDCap hosts. Run dynamic tests on cloned VMs with synthetic survey hashes.
Long-term hardening beyond the patch
After upgrading to 16.0.49 / 17.3.10 / 17.4.4, keep public surveys disabled unless a study protocol requires them. Place REDCap behind VPN or Zero Trust even when patches land—survey hashes will always be shareable URLs.
For agent-assisted survey design, never let models paste production survey links into public GitHub issues or Slack—that publishes the precondition Securifera describes as discoverable.
WAF rules should flag __passthru encodings and DataImportController in Referer paths originating from /surveys/ routes. Tune false positives in staging; do not wait for threat intel signatures after KEV.
Linking to clinical compliance narratives
Researchers searching “REDCap remote code execution” need plain-language next steps: patch, rotate credentials, notify IRB/security if PHI may have left the host. Beam does not store clinical data; it can document agent commands run against REDCap servers during investigations if hooks were enabled on admin laptops.
Timeline from disclosure to KEV
| Milestone | Date |
|---|---|
| Securifera publication | 2026-09-27 |
| GHSA / NVD | 2026-09-20 publication dates on advisory records |
| CISA KEV | 2026-11-18 |
The two-month gap explains why some universities patched in October while others learn urgency only from November KEV. If your change advisory closed in September, reopen based on exploitation, not novelty.
Defensive monitoring queries (conceptual)
Security teams without REDCap source access can still hunt:
- HTTP POST to paths containing
DataImportfrom Referer URLs withsurveyssegments. - Query strings with
__passthruusing percent-encoded slashes or double encoding. fnameparameters containingphp:,zip:,phar:, orconvert.base64filter chains Securifera describes.- Filesystem mtime changes under
application/cacheor Blade-equivalent temp directories during survey-only traffic.
Tune alerts to study production hours to reduce noise, but do not disable during exam weeks when public surveys peak.
Data handling after suspected compromise
Treat participant IDs, free-text responses, and imported CSVs as exfiltrated until disproven. Notify legal/privacy per institutional policy, separate from IT ticket closure. Agents helping draft notifications must not paste real PHI into cloud model prompts—use redaction patterns and local tooling only.
Vanderbilt product governance
REDCap is centrally versioned but locally hosted—each institution’s VM may lag. IT, research IT, and departmental admins each think the other patched. Assign a single version owner with monthly attestation posted to an internal wiki agents can read (minus secrets).
For multi-tenant hosting vendors, demand written confirmation of ≥ patched build on your instance, not only “REDCap cloud is secure.”
Comparison to other PHP application KEVs
| CVE family | Axis similarity |
|---|---|
| PHP2Shell / agent discovery | Also language-runtime issues adjacent to agent patching |
| Active Storage ImageMagick | File import primitives trust client paths |
REDCap’s chain is more auth-bypass-like despite no login: public survey is the entry role. Training materials for student RAs should say survey links are secrets—not passwords, but exploit preconditions.
Coordinating with November 2026 KEV cluster
The same CISA update that added REDCap added LiteLLM, Apache Axis, and Synology—different stacks, same patch weekend. Platform teams should publish a single internal status page listing owner, due date, and verification command per row so agent automation does not open six conflicting PRs that each claim emergency merge. Prioritize internet-facing research surveys above internal-only REDCap modules that never expose surveys/index.php links.
When coding agents generate IRB application text referencing REDCap version numbers, verify they quote live Control Center builds—stale version strings in PDFs mislead auditors into thinking 16.0.49 applied when 13.x still runs.
Export a weekly CSV of public survey URLs from production and compare against WAF allowlists—any orphan survey still live after a study closes is unused attack surface agents cannot see from git alone.
Schedule patch verification with a smoke-test survey on a staging clone before production upgrade; Securifera’s chain touches import, cache, and passthrough together, so unit tests agents write rarely cover full HTTP routing end-to-end.
Frequently asked questions
What is CVE-2026-90817?
Unauthenticated RCE in REDCap survey passthrough → Data Import chain.
Is it in CISA KEV?
Yes, added 2026-11-18, due 2026-11-21, forensics required.
Do I need a REDCap login?
No—only a valid public survey hash.
What are patched versions?
16.0.49 LTS, 17.3.10 LTS, 17.4.4 Standard+.
Can I disable surveys temporarily?
Yes—Securifera and OpenCVE cite disable public surveys as containment.
Does Beam protect REDCap?
No—Beam is workstation agent monitoring; patch REDCap on the server.
Where is the primary write-up?
Securifera and GHSA-hg88-qg2r-c8mx.
Securifera and CISA KEV as of November 18, 2026.

