Skip to content

armgit / AiSOC

Public Imported from GitHub
0

Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API keys, MIT licensed. Ships an MCP server for Claude, Cursor and Continue.

11.9 KB
Raw Blame

v15.0.0 — draft security advisories

Thirteen defects, all fixed in v15.0.0 (2026-10-01). This file is the draft text for the GitHub Security Advisory of each, kept in the repository so the wording is reviewable in a diff rather than typed once into a web form.

How to read the severities

Every CVSS vector below is v3.1 and is scored against a default deployment of v14.0.0, not against a worst-case configuration. Where a defect's impact depends on something an operator chose, that is stated in the vector's scope rather than assumed in the operator's favour.

Three of these are deliberately scored lower than their description might suggest, and the reason is given in each: a cookie no code reads, a control that is defence in depth behind a guard that was itself correct, and a path reachable only by an already-authenticated principal.

None of these were reported by an external researcher. They were found by reading the code at the v14.0.0 tag. No exploitation is known, and the disclosure is therefore retrospective rather than coordinated — the maintainer's decision, recorded in SECURITY_BATCH_PROGRESS.md as deviation D0.


S1. Anonymous requests were served as a tenant administrator

Severity: Critical — CVSS 3.1 9.1 AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:N CWE-287 Improper Authentication · CWE-1188 Insecure Default

is_dev_mode() granted an unauthenticated caller the admin role in the demo tenant, and DEMO_TENANT_ID was the same UUID that bootstrap_admin.py seeds as the first real tenant (00000000-0000-0000-0000-000000000001). A deployment that never set ENVIRONMENT therefore served anonymous requests as an administrator of its own production tenant.

Scope is Changed because the anonymous principal crosses into a tenant it was never issued for.

Fixed by requiring three independent conditions — ENVIRONMENT in the development set, an explicit AISOC_DEV_AUTH_BYPASS=1, and every published bind address being loopback — and by moving the demo tenant to a distinct UUID so the collision cannot recur.

Workaround for v14.0.0: set ENVIRONMENT=production.


S2. One tenant's Copilot conversations and saved hunts were readable by every other

Severity: High — CVSS 3.1 7.7 AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:L/A:N CWE-639 Authorization Bypass Through User-Controlled Key · CWE-668 Exposure to Wrong Sphere

Conversations and saved searches lived in module-level dictionaries with no tenant column. GET /api/v1/copilot/conversations returned every tenant's conversations to any authenticated caller; GET /api/v1/copilot/conversations/{id} returned any conversation to anyone holding its id.

This is not chat history. A conversation carries the analyst's question — naming hosts and users — and the model's answer, which quotes the alert evidence it was grounded on.

Integrity is Low rather than None because the same store accepted writes; Availability is None because a restart lost the data for everyone equally.

Fixed by migration 076, which moves both into tenant-scoped tables with tenant_id NOT NULL, forced RLS, and a query-layer predicate.


S3 / S4. Fabricated records served as tenant data

Severity: Low — CVSS 3.1 4.3 AV:N/AC:L/PR:L/UI:N/S:U/C:L/I:L/A:N CWE-1230 Exposure of Sensitive Information Through Metadata

/api/v1/shifts returned three invented shifts with named analysts and a fabricated ticket id; /api/v1/threatintel/stix/* returned invented indicators and collections. Both accepted writes into the same shared lists, so one tenant's POST was readable by another.

Scored Low deliberately: the data was fictional, so the confidentiality impact is the cross-tenant leak of what a caller wrote, not of real security telemetry. The more serious problem was honesty rather than disclosure.

Fixed by deleting the shift board (no caller existed anywhere) and answering 404 on the STIX reads outside demo mode. Breaking.


S5. No rate limit on credential verification

Severity: High — CVSS 3.1 7.5 AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N CWE-307 Improper Restriction of Excessive Authentication Attempts

POST /api/v1/auth/login applied no throttle, lockout or backoff, so password guessing was bounded only by network throughput.

Fixed by a two-counter throttle — per account and per source address — with exponential backoff and temporary lockout. A successful login clears the account counter but not the source counter, so a spray that happens to succeed once does not reset the attacker's budget.


S6. Both SSO handlers minted a session for an identity nobody authenticated

Severity: High — CVSS 3.1 8.1 AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:N CWE-287 Improper Authentication · CWE-1188 Insecure Default

app/auth/saml.py issued a signed token for stub-saml-user whenever python3-saml failed to import — and python3-saml was declared in no install path, so that branch was the only reachable path through the assertion consumer on every deployment. app/auth/oidc.py did the same for oidc-stub-user whenever OIDC_ISSUER was unset, which is the default.

Attack Complexity is High and this is not scored 9.8 for one reason: the API verifies a bearer token and does not read the aisoc_token cookie these set. The token was real and signed; nothing consumed it. A cookie the API ignores today is one it might read tomorrow, which is why this was fixed before SSO was made to work rather than after.

Fixed by deleting both branches, answering 501, and declaring python3-saml so the 501 is clearable.


S7. The audit log could not name the credential that acted

Severity: Medium — CVSS 3.1 5.3 AV:N/AC:L/PR:L/UI:N/S:U/C:L/I:N/A:N CWE-778 Insufficient Logging

An API key owned by a user resolved to that user's email, so an entry read identically whether the person acted at the console or a key they minted a year ago acted from a script. Revoking a key and disabling a person are different responses, and the log could not say which applied.

Separately, tenant_admin did not hold audit_log:read: the only roles that did held * across every tenant, so a customer could not review their own trail. That is a compliance failure against SOC 2 CC7.2 and ISO 27001 A.12.4 rather than a vulnerability, and is not scored.

Fixed by carrying the key prefix on the principal and recording auth_method on every entry, and by granting tenant_admin a tenant-scoped read.


S8. Six of seven mssp_* tables had no row-level security

Severity: Medium — CVSS 3.1 6.5 AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:N/A:N CWE-566 Authorization Bypass Through User-Controlled SQL Primary Key

Only mssp_tenant_metrics carried a policy. On the other six the query predicate was the single control between one MSSP's portfolio and another's.

Scored 6.5 rather than higher because this is defence in depth that was absent, not a reachable bypass: the route guards were correct at the time of the release. The reason it matters is the precedent — POST /mssp/overrides previously wrote a caller-supplied child tenant id onto a row the resolver read back filtered on the victim's tenant, and when that guard was wrong there was no second layer.

Fixed by migration 077. Forced RLS now covers 123 tables, up from 117.


S9. State-changing routes that authenticated and authorized nothing

Severity: Medium — CVSS 3.1 6.5 AV:N/AC:L/PR:L/UI:N/S:U/C:L/I:H/A:N CWE-862 Missing Authorization

POST /knowledge-base/query reached the tenant's knowledge base and could spend an LLM call on bare identity. Four MSSP routes managing portfolio membership were correctly guarded by an organisation-role dependency the coverage gate did not recognise — reported here for completeness because the published figure was wrong, not because those routes were open.

Fixed by a named knowledge_base:read entitlement, by teaching the gate to recognise organisation-role decisions, and by requiring every remaining identity-only route to carry a written reason.


S10. An AI triage could suppress an attacker's own alerts

Severity: High — CVSS 3.1 7.1 AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:H/A:L CWE-349 Acceptance of Extraneous Untrusted Data · CWE-807 Reliance on Untrusted Inputs in a Security Decision

Three AI-authored outcomes at ≥0.90 confidence on one evidence signature auto-closed every later alert sharing it, with no human having looked. The corroboration threshold was never a mitigation, because the attacker chooses how many alerts to send — raising it from three to thirty changes the cost, not the outcome.

Integrity is High because the result is a detection the SOC no longer sees; Availability is Low because the alert path remains functional for everything else. Confidentiality is None.

Fixed by restricting suppression to human-confirmed priors, expiring priors 90 days after their last confirmation, and refusing to suppress on evidence that tripped the prompt-injection guard.


S11. Every Kafka client, and every Postgres client, spoke cleartext

Severity: Medium — CVSS 3.1 6.8 AV:A/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:N CWE-319 Cleartext Transmission of Sensitive Information

Thirteen Kafka clients across Python, Go and TypeScript each took the library default, which is PLAINTEXT in all three. The spine carries raw event bodies, usernames, hostnames and command lines. sslmode was unset on every Postgres client, which libpq reads as prefer — TLS when offered, cleartext when not, with no way to tell which happened.

Attack Vector is Adjacent and Complexity High because this requires a position on the broker's network path.

Fixed by one transport resolver per language that refuses cleartext in a protected environment, and by reporting both an explicit and an unset sslmode at boot.


S12. A shipped integration credential

Severity: Medium — CVSS 3.1 5.9 AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:H/A:L CWE-1392 Use of Default Credentials

caldera_api_key defaulted to ADMIN123, Caldera's own published first-run credential. A deployment that never configured Caldera still built a client holding it and pointed it at whatever caldera_url resolved to — so the failure mode was not "purple-team does not work", it was "purple-team authenticates to a Caldera instance with the default password", which succeeds against any Caldera nobody rotated.

Fixed by removing the default and refusing with a 503 that names CALDERA_API_KEY.


S13. A permission granted in the console governed 27 routes out of 302

Severity: Medium — CVSS 3.1 6.5 AV:N/AC:L/PR:L/UI:N/S:U/C:L/I:H/A:N CWE-863 Incorrect Authorization

275 route dependencies read a hardcoded role map while 27 read the RBAC tables the console's administration screen writes to. An operator could grant a permission, watch it appear in the UI, and have 275 of 302 routes ignore it — and revoking worked no better, which is the direction that makes this a vulnerability rather than a usability defect.

Worse, the database path fell back to the static map whenever a user had no rows in user_roles, so removing every role from a user restored their static permissions.

Fixed by resolving permissions once at authentication from the database, and by distinguishing "this tenant has no RBAC configured" from "this user has been deprovisioned" — a question answered at the tenant level, not the user's row count.


What is deliberately not claimed

These severities are the maintainer's own assessment and have not been reviewed by a third party or assigned CVE identifiers. No exploitation of any of them is known, and no telemetry exists that would show it if it had occurred — which is itself worth stating, because "no evidence of exploitation" and "evidence of no exploitation" are different sentences and only the first is true here.