Research date: 2026-11-06. This document records what goraz learned from earlier projects. It is a summary of published material, not an assessment of those projects' overall quality.
How to read the evidence
A linked claim comes from a page that was read during this research.
(secondary) means the claim comes from a third-party article or a search excerpt, not from the project or the original researcher.
(unconfirmed) means no page that was read supports the claim. Do not cite it without checking.
"None found" means only that the searches found nothing. It does not show that nothing exists.
CVE scores and fixed versions vary between sources. Check the original advisory before relying on one.
Storage and servers must be treated as hostile. Researchers presented 12 attacks on Bitwarden, 7 on LastPass, 6 on Dashlane and 6 on 1Password under a malicious server. Causes include unauthenticated metadata and KDF settings, missing key separation, and legacy code paths.
Typed audit events with no free-form field, plus canary tests in CI
5
Revocation and sealing are often weaker than they sound.
Almond audit found Vault keys in memory after sealing; Infisical may accept a revoked token for up to 12 minutes during a Redis outage; OpenBao AppRole secret IDs outlived expiry until tidy ran
Check expiry and revocation at each use, fail closed, and state the limits plainly
Audit:Almond for ANSSI spent 60 person-days on 1.21.0 and found one vulnerability, 12 minor technical facts and 18 recommendations. Keys remained in memory after sealing (TF 6). Tokens sit in plaintext in ~/.vault-token (REC 2). Audit is off by default (REC 6).
Incidents and lessons:
Project Zero 2020: lenient identity-provider response parsing. Lesson: integrations are hostile parsers.
CVE-2025-6004 and CVE-2025-6014: case and trailing whitespace defeated lockout and TOTP replay protection. Lesson: canonicalize before any check.
CVE-2025-6011: a malformed dummy hash exposed timing. Lesson: make the failure path do the same work.
CVE-2026-5807: unauthenticated control endpoints caused denial of service.
Note: no breach of the Vault product itself was found in the pages read. Misconfigured deployments were not researched.
Lesson: use constant-time comparison on break-glass paths, and never interpolate user data into policy patterns. A fork inherits history and needs its own review.
Threat model: the security page orders risks by likelihood, from compromised cloud credentials to weak AES. It does not discuss integrity.
Integrity detail, confirmed in sops.go: the MAC is SHA-512 over the walked leaf values, written with no length framing. Key names and recipient metadata are not written to that hash within Tree.Encrypt.
Incidents:
GHSA-jgf3-f6rg-8x3h (High, 7.4): decrypting an untrusted file sends the user's Vault token to the address named in that file. Lesson: file metadata must never select a network endpoint that receives a credential.
GHSA-x4cm-pcq4-39j4: exec-file --filename ../x writes plaintext outside the temporary directory.
Issue #2243: a refactor changed MAC input and broke existing files. Lesson: authenticate exact bytes, and freeze formats with golden vectors.
Incidents:GHSA-32gq-x56h-299c: a plugin name with a path separator could execute an arbitrary binary. Lesson: identifiers that become exec targets need an allowlist.
Audit: no formal audit of age found, so goraz will not imply one by using it.
Incidents:CVE-2018-12356: unanchored parsing of GPG output allowed signature spoofing. CVE-2020-28086 (secondary): content was not bound to file name. Lessons: do not parse tool output for security decisions, and bind content to its name.
Incidents:GHSA-rw43-4vhf-hcvj and GHSA-v4c2-4fvx-chr3: a collaborator's crafted names and symlinks caused writes or deletes outside the store. Lesson: a store full of per-secret files inherits this whole class of bug.
Sealed Secrets: the clearest explanation of key renewal versus secret rotation.
Passbolt: published Cure53 results, including a High finding that clients do not verify metadata signatures. Lesson: signing without verifying gives no integrity.
Akeyless: novel fragment cryptography. Lesson: new constructions need independent review.
Palant 2023: server-side iterations do not protect the encryption key.
ETH Zurich: KDF removal gave a 300,000 times cracking speed-up. A ciphertext swapped into the URL field was sent to the icon server. An unprotected type prefix allowed a downgrade to CBC.
Vendor response: ten issues, seven resolved or in remediation, three accepted by design.
CLI 2026.4.0 on npm, CVE-2026-42994: malicious package between 5:57 and 7:30 PM ET on 2026-04-22. Bitwarden reports no evidence that vault data was accessed. A forum comment gives a slightly earlier publish time.
Threat model: vault cryptography lives in the Bitwarden clients.
Record: the October 2026 advisories include a revoked member keeping organisation access, and 2FA remember tokens that survive an administrator removing 2FA.
Lesson: end-to-end encryption does not make server authorization irrelevant.
Threat model: the audit excludes an untrustworthy OS and side channels. A sync provider can still roll the file back.
Audit:no major problems in core cryptography. The auditor recommended Argon2id and higher defaults. TOTP, browser integration, Auto-Type and HIBP were not reviewed.
Incidents:2026 advisories include memory bugs in the legacy .kdb importer and in browser integration. Lesson: importers and IPC are where bugs land.
KeePass (not XC):CVE-2023-32784 recovered the master password from memory dumps. CVE-2023-24055 (disputed) exported cleartext through a config trigger. Lessons: immutable strings leave residue, and config must not be able to weaken protection without authentication.
Threat model:Appendix A states that a malicious server could perform a man-in-the-middle attack on public keys, that removal does not change keys, and that malware can read an unlocked client.
runtime/secret requires GOEXPERIMENT=runtimesecret, carries no compatibility promise, and does nothing outside Linux amd64 and arm64 in the released documentation.
x/crypto/argon2 allocates its matrix on the ordinary heap, has no wiping code and panics on invalid parameters. RFC 9106 advises memory wiping where side channels are a concern, so this is a known gap.
memguard offers off-heap locked memory, but key bytes still enter the Go heap when passed to standard cipher packages.
NVD pages rendered empty, so several CVE details come from project advisories or aggregators.
Cyata's and CyberArk's own write-ups on Vault Fault could not be opened, so that chain rests on secondary reporting.
OpenBao advisory list pages 4 and 6 were not read.
The Cure53 reports on Bitwarden were not read in full.
No formal threat model was available for Doppler, Conjur or Teller, and no audit report was available for many others. A "None found" entry should not be read as a criticism.