Hardware security modules are very good at one thing: keeping a private key from leaving a protected boundary. They are not, on their own, a statement about who may ask that key to sign. Some of the best-documented signing failures of recent years were not key theft. They were cases where a legitimate signing capability was used for something it should not have been. This article sets out what primary sources say about three such cases, then how we think about signing authority as a control in its own right.
How to read this article. “Established facts” restates what named primary sources say, with numbered references. “SafeCipher analysis” is our own opinion. The cases below are real and sourced. Where we discuss AI agents, we cite no incident, because we have found no well-documented public case from a primary source, and we say so.
Established facts
Three documented cases
A signing service signed malicious code. On 25 June 2021 Microsoft reported that an actor had submitted drivers for certification through the Windows Hardware Compatibility Program and that the drivers were malicious. Microsoft said it had suspended the account and reviewed the actor’s other submissions, and stated: “We have seen no evidence that the WHCP signing certificate was exposed. The infrastructure was not compromised.” It also said the driver’s use occurs after an attacker already has administrative privileges or persuades a user to install it, and that it would refine partner access policies, validation and the signing process.1 The signature was valid because the signing process had accepted the submission.
A build pipeline’s output was signed. CISA’s alert on the SolarWinds Orion compromise states that the adversary added a malicious version of a component into the SolarWinds software lifecycle, “which was then signed by the legitimate SolarWinds code signing certificate”.2 The key was used as designed; the thing being signed was not what the signer intended.
Pipeline secrets were stolen. CircleCI’s incident report describes an unauthorised party stealing a valid, two-factor-backed single sign-on session from an engineer’s laptop. The engineer had privileges to generate production access tokens, and the party exfiltrated customer environment variables, tokens and keys. CircleCI says the data was encrypted at rest, but the party extracted encryption keys from a running process. It recommended customers use OIDC tokens to avoid storing long-lived credentials, and described planned moves toward more ephemeral permissions.3
What standards say about the key, and what they leave open
- For code-signing certificates, the CA/Browser Forum Baseline Requirements (version 3.11.0, 16 June 2026) say that, effective 1 June 2023, subscribers must generate and protect private keys in a hardware crypto module certified to at least FIPS 140-2 Level 2 or Common Criteria EAL 4+. The permitted routes include a cloud-based key generation and protection solution, which must log all access, operations and configuration changes, or a signing service that meets the requirements of section 6.2.7.3.4 That protects the key and requires logging. It does not say who may request signatures or on what approval.
- NIST SP 800-57 Part 1 Revision 5 provides general guidance and best practices for the management of cryptographic keying material, including the functions involved in key management.5 NIST SP 800-130 is a framework of topics a designer should address in a key management system design specification, and SP 800-152 is a profile of requirements for US federal key management systems based on it.67 NIST’s catalogue lists SP 800-57 Part 1 Revision 6 as a draft released on 5 December 2025.8
SafeCipher analysis
The points below are our engineering view, not a statement of what a standard requires.
The question to ask is “who or what can cause this key to sign, and under what conditions?” The HSM answers where the key lives. Signing authority is the set of identities, permissions, approvals and records around the key. In each documented case above, the cryptography worked. The control that failed or was bypassed was about authority: what was submitted, what was built, or who held access to the means of obtaining signatures.
We find it useful to split signing authority into five parts.
1. Identity. Every signing request should come from a named, authenticated principal: a person, a service or a workload. Shared credentials and long-lived tokens in pipeline settings make this impossible to establish after the fact. Short-lived, workload-bound credentials (the OIDC approach CircleCI recommends3) are a better basis than static secrets.
2. Authorisation. Decide in advance which principal may use which key, for which kind of artefact, from which environment. A key that signs production releases should not be reachable from a developer’s workstation, and a test key should not be able to produce a production-trusted signature. Where products support it, bind permissions to the HSM or KMS key policy rather than only to the pipeline.
3. Approval. Authorisation says a request is allowed in principle. Approval says this particular request is acceptable. For high-impact keys, such as release signing, CA keys or code-signing keys, require a second party or a quorum, and make the approver see what will be signed (the artefact digest, source revision and build provenance), not just a request ID. Approval of an opaque request is a rubber stamp.
4. Audit evidence. Keep a record that ties together who requested, what artefact digest was signed, which key, who approved, and the HSM or KMS log entry. Reconcile that record against what was actually released. If you cannot answer “list every signature produced with this key last quarter and its source build,” you cannot detect misuse after the event.
5. Revocation of authority. Key revocation is blunt. Plan for the more common event: revoking a person’s, service’s or pipeline’s right to sign without rotating the key. That means being able to disable roles and credentials quickly, expire grants, and know which artefacts were signed in the affected window, and what timestamping and revocation implications follow.
Where CI/CD, administrators and AI agents fit
CI/CD. Treat the pipeline as the signing principal and the signing service as the control point. A sound pattern is a dedicated signing service that signs only artefacts with verified provenance from a controlled build, rather than giving pipeline jobs direct access to key operations. Avoid long-lived signing credentials in build configuration.
Administrators. HSM and KMS administrators can often change policies and grant roles. Separate the ability to administer from the ability to sign, record break-glass use, and review administrative actions as carefully as signatures.
AI agents. [SafeCipher analysis, not an incident report] An autonomous agent that can call a signing or key-management API is another principal with authority, and an unusual one: its requests are shaped by prompts and tool outputs that an attacker may influence. We cite no documented incident of this kind from a primary source. The risk is a design concern, not a reported event. Our recommendations are to avoid giving agents standing permission to perform sign, decrypt or key-administration operations; to scope any permission to a single action type and key; to require human or quorum approval of high-impact operations with the artefact visible to the approver; to rate-limit and log every call with the agent’s identity and the task context; and to make it fast to revoke the agent’s credentials. This is the same set of controls as above, applied to a principal that is easier to manipulate than most.
Reading the three cases through the five parts
- WHCP drivers (Microsoft). The key was not exposed, according to Microsoft. The weak point was the decision to accept and sign a submission: an authorisation and approval question, and an audit question after the event.
- SolarWinds. A build-process compromise meant that the signing step certified an artefact the vendor did not intend. Identity of the requester was fine; the evidence linking “what was built” to “what was signed” was what mattered.
- CircleCI. A stolen session and standing production privileges gave access to secrets. The relevant control is limiting standing privilege and credential lifetime, and being able to revoke quickly, which CircleCI itself describes as its direction.3
We do not suggest that any of these organisations lacked HSMs or controls. The point is only that a hardware boundary around the key answers a different question from “who may cause a signature”.
Questions to put to your HSM, KMS or signing-service supplier
- Can we separate roles for administering, approving and using a key, and are those roles enforced by the product rather than by convention?
- Can high-impact operations require more than one person, and what does the approver see before they approve?
- What does the log record for each signing call (caller identity, key, data digest, time), and can it be exported to our own tamper-evident store?
- Can we scope a credential to one key and one operation, and make it short-lived?
- How quickly can we disable one caller without affecting others or rotating the key?
- For cloud key management, which of these controls are configuration choices we must make, and which are provider defaults?
A signing-authority checklist
- ☐ Every key has a named owner and a documented purpose
- ☐ Every signing request is tied to an authenticated principal, with no shared or static signing credentials
- ☐ Permissions are scoped by key, artefact type and environment
- ☐ High-impact keys require a second party or quorum who sees the artefact digest and provenance
- ☐ Signing records link requester, artefact digest, key, approver and HSM/KMS log entry
- ☐ Signatures can be reconciled against actual releases on a regular schedule
- ☐ Authority can be revoked quickly without rotating the key
- ☐ Administration is separated from signing, and break-glass use is recorded
- ☐ AI agents and automation have no standing signing or key-administration rights
- ☐ Key custody (HSM, cloud key protection or signing service) is reviewed alongside authority
What this article does not claim
We have not examined the internal controls of Microsoft, SolarWinds or CircleCI beyond what their own and government statements say. We make no allegation about any organisation. The three cases show different failures, and we use them only to illustrate that key protection and signing authority are separate questions.
Next step
SafeCipher reviews key custody and signing authority across HSMs, cloud key management and signing pipelines, independently of any manufacturer. If you want to know who and what can sign in your environment, start with a cryptographic audit, or read about our HSM and key management architecture and vendor-neutral HSM comparison. For the wider governance picture see The Cryptographic Void: XIoT, Agentic AI, and the Erosion of Authority.
References
All sources accessed on 16 November 2026.
- Microsoft Security Response Center, “Investigating and Mitigating Malicious Drivers”, 25 Jun 2021: microsoft.com/en-us/msrc/blog/2021/06/investigating-and-mitigating-malicious-drivers
- CISA, Alert AA20-352A, “Advanced Persistent Threat Compromise of Government Agencies, Critical Infrastructure, and Private Sector Organizations”: cisa.gov/news-events/cybersecurity-advisories/aa20-352a
- CircleCI, “CircleCI Jan 4, 2023 security incident report”, 12 Jan 2023: circleci.com/blog/jan-4-2023-incident-report
- CA/Browser Forum, Baseline Requirements for the Issuance and Management of Publicly-Trusted Code Signing Certificates, version 3.11.0 (16 June 2026), section 6.2.7.4.1: cabforum.org/uploads/CA-Browser-Forum-CSCBR-3.11.0.pdf
- NIST SP 800-57 Part 1 Rev. 5, Recommendation for Key Management: Part 1 General (May 2020): csrc.nist.gov/pubs/sp/800/57/pt1/r5/final
- NIST SP 800-130, A Framework for Designing Cryptographic Key Management Systems (Aug 2013): csrc.nist.gov/pubs/sp/800/130/final
- NIST SP 800-152, A Profile for U.S. Federal Cryptographic Key Management Systems (Oct 2015): nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-152.pdf
- NIST CSRC, SP 800-57 Part 1 Rev. 6 (Initial Public Draft, 5 Dec 2025): csrc.nist.gov/pubs/sp/800/57/pt1/r6/ipd

Leave a Reply