Best Secrets Managers for CI/CD in 2026: 7 Compared After wfx-loot
Last week's wfx-loot campaign was a clean demonstration of what a stored CI secret is worth. The injected step did one thing: it passed toJSON(secrets) to a shell script. Every repository secret, every environment secret the job could reach, gone in one run. Then it used the job's OIDC token to mint fresh cloud credentials, which is the part people miss. Moving secrets out of GitHub is only half the fix. The other half is making sure a stolen or hijacked CI identity can only reach what one specific workflow, on one specific branch, in one specific environment, needs.
That second half is what this comparison ranks. I compared eight products, counting Vault and OpenBao as one entry because OpenBao is a fork that keeps Vault's API. All of them can hold a secret. They differ in how a pipeline logs in, how tightly you can bind that login, whether they issue short-lived credentials, and what they cost.
Method, and what "tested" means here
- Criteria, in order of weight: (1) can a CI job log in with its OIDC token instead of a stored token; (2) can the role be bound to repository, branch and environment claims; (3) can it issue dynamic, short-lived credentials; (4) licence and self-hosting; (5) price at a stated workload.
- Hands-on: I ran HashiCorp Vault locally in dev mode and tested JWT login with tokens that carry GitHub-style claims, including the branch names wfx-loot used. Results are below. I did not run OpenBao, Infisical, Doppler, 1Password or the cloud services for this piece. For those, every claim comes from the vendor's own docs, pricing page or GitHub repository, read on 16 November 2026.
- Versions: latest GitHub releases as of 16 November: Vault v2.2.2 (12 November), OpenBao v2.7.2 (26 October), Infisical v0.167.10 (13 November).
- No paid placements. Nobody paid to be here and there are no affiliate links. I build PandaStack, which is not a secrets manager and is not in this list.
- Last tested: 16 November 2026.
The comparison
| Product | CI login without a stored token | Bind to branch and environment | Dynamic secrets | Licence or model | Entry price |
|---|---|---|---|---|---|
| HashiCorp Vault 2.2 | JWT auth with GitHub's OIDC issuer | Yes, bound_claims on any claim, globs optional (docs) |
Yes, in the open edition | BUSL 1.1 from 1.15.0, converts to MPL 2.0 after four years (LICENSE) | Free to self-host for internal use |
| OpenBao 2.7 | JWT/OIDC auth method, same API as Vault (docs) | Yes, bound_claims |
Yes | MPL 2.0; OpenSSF Sandbox project under the Linux Foundation (site) | Free, self-host only |
| Infisical | OIDC auth with a GitHub guide (docs) | Yes, Subject, Audiences and Claims fields | Advanced plan and up | MIT except the ee directories (repo) |
Free for 5 identities; Pro $20 per identity per month annual (pricing) |
| Doppler | Service account identities over OIDC, Team plan and up (docs) | Yes, audience and subject required, custom claims optional | Enterprise plan | SaaS | Free for 3 users; Team $21 per user per month (pricing) |
| 1Password | Workload Identity in load-secrets-action, public preview (repo) |
Not documented in the action README | No | SaaS; action is MIT | Not checked |
| AWS Secrets Manager | Through IAM: GitHub OIDC to an IAM role | Yes, in the role's trust policy | Rotation functions, not per-job credentials | Managed service | $0.40 per secret per month plus $0.05 per 10,000 calls (pricing) |
| Google Secret Manager | Through Workload Identity Federation | Yes, in the provider's attribute condition | No | Managed service | $0.06 per active version per month, 6 free; $0.03 per 10,000 accesses, 10,000 free (pricing) |
| Azure Key Vault | Through a Microsoft Entra federated credential | Yes, the credential matches an exact subject | No | Managed service | $0.03 per 10,000 operations, Standard or Premium (pricing, read from the Azure retail prices API) |
Two notes on that table. Infisical counts "any principal that authenticates to Infisical, human or machine" as an identity, so every CI machine identity is a billable seat on the paid plans. Doppler goes the other way: its pricing page says "AI agents and non-human identities ride free", and you pay for users.
What I tested: a Vault role that refuses the attacker's branches
The question I wanted answered is narrow: if an attacker gets a commit into your repository, as wfx-loot did, which of their runs can still log in? I used the Vault binary already on this Mac, which is 1.18.3, older than the current 2.2.2. The JWT auth method's claim checks are the same feature in both, but I did not re-run on 2.2.2. I could not get GitHub to sign tokens for a test repository that does not exist, so I generated an RSA key, gave Vault the public half with jwt_validation_pubkeys, and signed tokens myself with GitHub's issuer string and the claims GitHub documents. That tests Vault's claim matching, not GitHub's token service.
export VAULT_ADDR=http://127.0.0.1:18200 VAULT_TOKEN=root
vault server -dev -dev-root-token-id=root -dev-listen-address=127.0.0.1:18200 &
vault kv put secret/ci/deploy api_key=test-not-real
vault policy write ci-deploy - <<'EOF'
path "secret/data/ci/deploy" { capabilities = ["read"] }
EOF
vault auth enable jwt
vault write auth/jwt/config \
jwt_validation_pubkeys=@pub.pem \
bound_issuer=https://token.actions.githubusercontent.com
The role is written as one JSON document, because bound_claims is a map and the CLI cannot set map values one key at a time, a limitation the OpenBao docs also point out:
{
"role_type": "jwt",
"user_claim": "repository",
"bound_audiences": ["vault"],
"bound_claims": {
"repository": "acme/app",
"ref": "refs/heads/main",
"environment": "prod"
},
"token_policies": ["ci-deploy"],
"token_ttl": "5m",
"token_max_ttl": "10m"
}
Then five kinds of login, each with a token signed by the same key:
| Token claims | Result | Vault's error |
|---|---|---|
acme/app, refs/heads/main, environment prod |
Accepted, secret read | None |
Same, but ref: refs/heads/chore/k3x9a2qz |
Rejected | claim "ref" does not match any associated bound claim values |
ref: refs/heads/wfx-docs, no environment claim |
Rejected | claim "environment" is missing |
repository: evil/fork |
Rejected | claim "repository" does not match any associated bound claim values |
| Valid claims, expired two minutes ago | Rejected | token is expired |
I ran the accepted case six times in all, and it succeeded every time. Each login took about 200 ms end to end, but that is almost all Python and Vault CLI start-up on my laptop, so I am not treating it as a performance number.
What this shows, and what it does not. Binding to ref stops the wfx-loot variant that pushed a new workflow on a throwaway chore/ branch. Requiring an environment claim stops a probe pull request or any job that does not target the environment. Neither stops the variant that added steps to an existing workflow on main that already deploys to prod. That run carries exactly the claims the role expects. The control that catches it is outside the secrets manager: required reviewers on the GitHub environment, so the job waits for a human before GitHub issues a token with environment: prod, and the reviewer is not the hijacked account. Keep token_ttl short as well, so whatever a run does get stops working within minutes.
The picks
Best self-hosted, and my default: OpenBao. It has Vault's JWT auth, its bound_claims and its dynamic secrets engines, under MPL 2.0 with Linux Foundation governance. If you are starting fresh and will run it yourself, there is little reason to accept a source-available licence. The cost is operational: you run an HA cluster, unseal, back up and upgrade it. My External Secrets Operator with Vault guide works against either.
Best if you already run it: HashiCorp Vault. The feature set is still the broadest, and internal production use is allowed under BUSL 1.1. Moving an existing estate to OpenBao for licence reasons alone is a project with a real cost. Do it when you have a reason beyond the licence.
Best on a single cloud: the cloud's own secret store, and often no secret at all. For cloud credentials, the strongest answer to wfx-loot is not storing any. GitHub OIDC straight to an IAM role, a Workload Identity pool or an Entra federated credential gives the job short-lived cloud access bound to repository, branch and environment, with nothing in GitHub to dump. Use AWS Secrets Manager, Google Secret Manager or Azure Key Vault for the third-party secrets that remain. On price at 200 secrets and one million reads a month, using the list prices in the table: AWS comes to $85.00, Google $14.61 and Azure $3.00. AWS's new cross-account replication, announced on 13 November, also lets a CI account hold a replica encrypted with its own KMS key instead of reaching into production. The AWS pages I read do not say whether replicas are billed.
Best managed, open-source option: Infisical. OIDC login from GitHub with subject, audience and claim checks, a self-hostable MIT core, and a free tier of five identities. The catch is the per-identity bill on paid plans once machine identities pile up, and dynamic secrets start at the Advanced plan, $40 per identity per month annually.
Best for developer experience: Doppler. Machine identities are free and the OIDC login requires audience and subject checks by design. OIDC identities need the Team plan at $21 per user per month, and dynamic secrets are Enterprise only.
Fine if you are already a 1Password shop. Workload Identity in load-secrets-action exchanges the GitHub OIDC token "for short-lived 1Password access", but it is a public preview, and the README does not say how the exchange is bound to branch or environment. I would wait for that to be documented before relying on it for production deploy credentials.
GitHub Actions secrets themselves are not on the list because they are the baseline wfx-loot emptied. They are fine for low-value values. Anything that can deploy or spend money should come from one of the above at run time, through a role bound the way the test shows.
What to do this week
- List every secret in your GitHub organisation and repositories, and mark which ones can deploy, publish a package or spend money.
- Replace cloud keys with OIDC federation first. It removes the most valuable loot for the least work.
- For the rest, pick one manager from the list and bind every CI role to repository, branch and environment, with a token TTL of minutes, not hours.
- Add required reviewers to the environments those roles require.
FAQ
What is the best secrets manager for GitHub Actions in 2026?
For cloud credentials, none: use GitHub's OIDC federation with AWS, Google Cloud or Azure so nothing is stored. For other secrets, OpenBao or Vault if you self-host, Infisical or Doppler if you want a managed service, and your cloud's own secret store if you live on one cloud. Whatever you pick, bind the CI role to repository, branch and environment claims.
Is Vault still open source?
No. HashiCorp Vault 1.15.0 and later are under the Business Source License 1.1, which allows internal production use but not offering Vault as a competing hosted product. Each version converts to MPL 2.0 four years after release. OpenBao is the MPL 2.0 fork, run as an OpenSSF Sandbox project under the Linux Foundation.
Would a secrets manager have stopped wfx-loot?
Partly. With secrets fetched at run time through a role bound to main and a prod environment, the probe pull requests and the throwaway-branch workflows could not have logged in, as my Vault test shows. A step injected into a workflow on main that already deploys to prod would still log in. Required reviewers on the environment and short token lifetimes are what limit that case.
How much does AWS Secrets Manager cost compared with Google and Azure?
At list price, AWS charges $0.40 per secret per month plus $0.05 per 10,000 API calls, Google charges $0.06 per active secret version per month after six free plus $0.03 per 10,000 accesses after 10,000 free, and Azure Key Vault charges $0.03 per 10,000 operations with no per-secret fee. For 200 secrets read one million times a month that is about $85, $14.61 and $3.
Related: Kubernetes Secrets in GitOps: SOPS vs Sealed Secrets vs ESO, Workload Identity Federation Explained and GitHub Actions Security in 2026.
Spot an error or something out of date? Email me and I will fix it and log it on the corrections page.
I'm Ajay Kumar — I build and operate PandaStack, an open-source Firecracker microVM cloud for AI agents, and write about the infrastructure and tooling around it.
Need this kind of infrastructure work? See what I do or email hello@ajayk.sh.
Related
Renovate vs Dependabot in 2026: 119 Managers vs 34, Both Wait 3 Days
Renovate vs Dependabot in 2026, compared on real defaults: 119 Renovate managers vs 34 Dependabot ecosystems, grouping, native auto-merge vs an Actions workflow, Dependabot's new 3-day cooldown, minimumReleaseAge, Mend-hosted vs self-hosted, and which to pick.
12 minNov 10, 2026SBOM Explained (2026): SPDX 3.0 vs CycloneDX 1.7, VEX and the Rules
What an SBOM is in 2026 and how to pick a format: SPDX 3.0.1 vs CycloneDX 1.7 (ECMA-424), CISA's 2026 minimum elements, the EU CRA's SBOM duty, VEX, why SBOM quality is poor, and the Syft, Trivy, cdxgen and Dependency-Track toolchain.
11 minOct 31, 2026SAST vs DAST vs SCA vs IAST in 2026: What Each Finds, 34% False Positives
SAST, DAST, SCA and IAST explained for 2026: what each scanner actually finds, real false positive rates (CodeQL at 34% in an ISSTA 2025 study), Semgrep vs CodeQL vs SonarQube vs Snyk vs ZAP, AI triage, and where each belongs in a CI/CD pipeline.
13 min