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.

chore(release): v12.1.0 (#1006)

Beenu Arora2a2b437436e7
diff --git a/.github/workflows/container-egress.yml b/.github/workflows/container-egress.yml
index 6904cc66f..ce97080c4 100644
--- a/.github/workflows/container-egress.yml
+++ b/.github/workflows/container-egress.yml
@@ -37,8 +37,14 @@ permissions:
contents: read
concurrency:
- group: container-egress-${{ github.ref }}
- cancel-in-progress: true
+ # Supersede a pull request's own earlier runs; never supersede a push.
+ # Keyed on the commit rather than the ref, because a group shared across
+ # commits cancels a run still pending when the next push queues — and a
+ # cancelled run reports `cancelled` rather than `failure`, so a commit
+ # nothing graded sits under a tick that does not look wrong.
+ # Gate: scripts/check_workflow_concurrency.py
+ group: ${{ github.workflow }}-${{ github.ref }}-${{ github.event_name }}-${{ github.event.pull_request.number || github.sha }}
+ cancel-in-progress: ${{ github.event_name == 'pull_request' }}
jobs:
container-egress:
diff --git a/.security/allowlist.yml b/.security/allowlist.yml
index aca7c032c..1219583bf 100644
--- a/.security/allowlist.yml
+++ b/.security/allowlist.yml
@@ -44,6 +44,19 @@ ceilings:
# argument to `urlopen`, and counting it is better than suppressing it.
# `entries:` below stays empty on purpose: a ceiling counts a finding, a
# suppression hides it.
+ #
+ # Unchanged at v12.1.0, and the reason is worth recording rather than
+ # smoothing over. Fixing GHSA-w754-prh8-m56j removed one ERROR
+ # (`use-defused-xml`, which fired on the PAN-OS client's
+ # `xml.sax.saxutils` import — string escaping with no parser and no XXE
+ # surface, which the rule cannot distinguish from parsing; replaced with
+ # `html.escape`). CI measured 39/101 at that point and asked for the
+ # ceiling to be lowered. Adding the rest of the release commit took the
+ # measurement back to 40/102, and **I could not attribute which file
+ # contributed the returning finding** — the scan covered the same 2,265
+ # files both times and `main` had not moved. Left at the measured value
+ # rather than lowered on a number that then moved under it; the honest
+ # state is that the net is unchanged and one finding is unexplained.
total: 102
error: 40
measured_with: "1.86.0"
diff --git a/CHANGELOG.md b/CHANGELOG.md
index ec55ad940..f2b04db53 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,40 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
## [Unreleased]
+## [12.1.0] - 2026-09-28
+
+### Security
+
+- **XML injection in the PAN-OS user-id client turned containment into its
+ opposite** ([GHSA-w754-prh8-m56j](https://github.com/beenuar/AiSOC/security/advisories/GHSA-w754-prh8-m56j)).
+ `_xml_register` and `_xml_unregister` interpolated `ip` and `tag` from
+ `ActionRequest.parameters` into a user-id message with f-strings, so a value
+ closing the `ip` attribute could append further payloads to the same message.
+
+ The injected payload is valid PAN-OS, which is what raises this above a
+ parsing curiosity: reproduced against the shipped code, `block_ip` emitted
+ well-formed XML containing an `unregister` for an address of the caller's
+ choosing, so the containment action released a block instead of applying one
+ and the firewall answered success.
+
+ Validated before escaped, in that order deliberately. Escaping alone stops
+ the injection and still hands the firewall a string that is not an address,
+ where it matches nothing and reports success — trading a loud failure for a
+ silent one on a containment path. Only the three forms PAN-OS registers are
+ accepted (single address, CIDR network, hyphenated range, with family and
+ ordering checked); tags are restricted to `[A-Za-z0-9._-]{1,127}`.
+
+- **The incident report's fallback renderer emitted untrusted HTML.** Found by
+ sweeping for the shape above. `_md_to_html` wraps the report in `<pre>` when
+ the `markdown` package is absent and interpolated the Markdown raw, plus the
+ case id raw into `<title>`. The sibling renderer in `orchestrator/report.py`
+ already escapes both, and its comment says why in as many words — one copy of
+ the lesson was written down and the other did the thing it warns against. The
+ content is model output and connector-supplied entity names.
+
+- **Eight advisories fixed in 12.0.0 are now published.** No code change; they
+ had remained in triage since the release that fixed them.
+
### Changed
- **Six of the eight `PARTIAL` rows in the claim-to-gate matrix closed, and
diff --git a/README.md b/README.md
index ed8096954..a5f578aa5 100644
--- a/README.md
+++ b/README.md
@@ -7,7 +7,7 @@
**An open-source, self-hostable AI Security Operations Center.** It ingests your security telemetry, detects and correlates threats, investigates them with AI agents whose reasoning is fully auditable, and proposes responses a human approves.
[![License: MIT](https://img.shields.io/badge/License-MIT-22c55e.svg?style=flat-square)](https://opensource.org/licenses/MIT)
-[![Version](https://img.shields.io/badge/version-12.0.0-f59e0b?style=flat-square)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-12.1.0-f59e0b?style=flat-square)](CHANGELOG.md)
[![CI](https://img.shields.io/github/actions/workflow/status/beenuar/AiSOC/ci.yml?branch=main&label=CI&style=flat-square)](https://github.com/beenuar/AiSOC/actions/workflows/ci.yml)
[![CodeQL](https://img.shields.io/github/actions/workflow/status/beenuar/AiSOC/codeql.yml?branch=main&label=CodeQL&style=flat-square)](https://github.com/beenuar/AiSOC/actions/workflows/codeql.yml)
[![OpenSSF Scorecard](https://api.securityscorecards.dev/projects/github.com/beenuar/AiSOC/badge)](https://securityscorecards.dev/viewer/?uri=github.com/beenuar/AiSOC)
diff --git a/RELEASES.md b/RELEASES.md
index 136eb44a3..b57d41e39 100644
--- a/RELEASES.md
+++ b/RELEASES.md
@@ -2,28 +2,50 @@
This file mirrors what used to live in the "What's new" section of [`README.md`](README.md). The complete, machine-readable inventory (with file paths, env-var diffs, and per-release test counts) lives in [`CHANGELOG.md`](CHANGELOG.md).
-> **TL;DR for first-time visitors:** AiSOC is on `v12.0.0`, released 2026-09-28 — the release where two services stopped serving unauthenticated and started refusing to serve instead. **Read the BREAKING section of the changelog before upgrading.** The actions service had skipped authentication entirely whenever its token was empty and `AISOC_DEV_MODE` was set, and compose defaults that flag on while nothing generated the token, so every stock install dispatched `isolate_host`, `disable_user`, `block_ip` and `run_script` to anything that could reach the port; the realtime edge verified connection tickets against a constant committed to this repository and treated an unset internal token as authorized; and a `viewer` role could write to cases on all nine write routes where the permission was enforced nowhere. Each was a reported vulnerability — four advisories across the three items — and each now returns 503 or 401 until the credential it needs is set, which is the intended failure and the reason this is a major. The upgrade action for all three is `make up`, because `scripts/ensure_env.py` backfills generated secrets into an **existing** `.env` and not only a new one. Alongside that, hunting stopped being a library of hunts somebody had already written: an agent turns a hypothesis into a plan whose fields and operators are closed enums, so there is no string it can produce that becomes SQL text, and the hunt corpus went from 5 to 68. Every product claim is backed by a failing CI test (claim-to-gate matrix at the v12.0.0 cut: 236 rows — 228 GATED / 8 PARTIAL / 0 NO GATE, counted by `scripts/check_claim_gate_matrix.py`, which is the figure to recount rather than to quote). The latest GitHub release with notes and downloads: <https://github.com/beenuar/AiSOC/releases/latest>.
+> **TL;DR for first-time visitors:** AiSOC is on `v12.1.0`, released 2026-09-28 — the release where two services stopped serving unauthenticated and started refusing to serve instead. **Read the BREAKING section of the changelog before upgrading.** The actions service had skipped authentication entirely whenever its token was empty and `AISOC_DEV_MODE` was set, and compose defaults that flag on while nothing generated the token, so every stock install dispatched `isolate_host`, `disable_user`, `block_ip` and `run_script` to anything that could reach the port; the realtime edge verified connection tickets against a constant committed to this repository and treated an unset internal token as authorized; and a `viewer` role could write to cases on all nine write routes where the permission was enforced nowhere. Each was a reported vulnerability — four advisories across the three items — and each now returns 503 or 401 until the credential it needs is set, which is the intended failure and the reason this is a major. The upgrade action for all three is `make up`, because `scripts/ensure_env.py` backfills generated secrets into an **existing** `.env` and not only a new one. Alongside that, hunting stopped being a library of hunts somebody had already written: an agent turns a hypothesis into a plan whose fields and operators are closed enums, so there is no string it can produce that becomes SQL text, and the hunt corpus went from 5 to 68. Every product claim is backed by a failing CI test (claim-to-gate matrix at the v12.0.0 cut: 236 rows — 228 GATED / 8 PARTIAL / 0 NO GATE, counted by `scripts/check_claim_gate_matrix.py`, which is the figure to recount rather than to quote). The latest GitHub release with notes and downloads: <https://github.com/beenuar/AiSOC/releases/latest>.
---
## What's new
-`VERSION` is `12.0.0`. The **v12.0.0** release (2026-09-28) treats the stack a stranger actually starts as the thing that has to be secure. Three of its items are reported vulnerabilities in the deployment this repository ships by default, and none of them needed a credential to reach.
+`VERSION` is `12.1.0`. The **v12.1.0** release (2026-09-28) is a security and
+correctness release. Its theme is narrower than v12.0.0's and worth stating
+plainly: several controls were reading a description of the system rather than
+the system, and reported success about the case they could not see.
+
+**v12.1.0 highlights (September 28, 2026)**
+- **A firewall containment action could be made to release a block.** The PAN-OS
+ user-id client interpolated the caller's `ip` and `tag` into XML with
+ f-strings, and the injected payload is valid PAN-OS — so `block_ip` emitted a
+ message telling the firewall to *unregister* an address of the attacker's
+ choosing, and the firewall answered success
+ ([GHSA-w754-prh8-m56j](https://github.com/beenuar/AiSOC/security/advisories/GHSA-w754-prh8-m56j)).
+- **The published API image answered 500 to every `/api/v1` request.** A tracing
+ middleware raised on the FastAPI version the lock resolved. CI installed the
+ same package with no version bound, resolved a newer one, and passed — so CI
+ and production were running different code, and the version CI never graded
+ was the broken one.
+- **Sigma rules were evaluated by substring containment**, which cannot express
+ a boolean, so every multi-clause rule matched nothing while reporting no
+ error. The published executable-rule count does not rest on that path and did
+ not change.
+- **Eight advisories fixed in v12.0.0 are now published.** They had remained in
+ triage since the release that fixed them.
+- **A marketplace catalogue of 7,155 items now says which are executable** —
+ 2,767 executable against 4,388 reference-only, tied to the detection truth
+ table rather than to a flag that the compiler does not rewrite.
+- **The audit hash chain can no longer fork.** Appends are serialized per tenant
+ and a unique index makes two rows claiming the same predecessor
+ unrepresentable; measured at 500 concurrent appends, 0 forks against 401
+ before.
+
+The full inventory — 16 entries — lives under `[12.1.0]` in [`CHANGELOG.md`](CHANGELOG.md).
-**Read the BREAKING section of the changelog before upgrading.** Two services now refuse to serve rather than serve unauthenticated, and one role loses a permission it was never meant to have. The actions service returns 503 on every mutating route until `AISOC_ACTIONS_SERVICE_TOKEN` is set; the realtime edge rejects every connection until `AISOC_REALTIME_JWT_SECRET` is set and answers 503 on `/internal/*` until `REALTIME_INTERNAL_TOKEN` is set. A deployment that upgrades without those values keeps running and stops responding on those routes, which is the intended failure and the reason this is a major. **The action for all three is `make up`** (or `make env`): `scripts/ensure_env.py` backfills every generated secret into an **existing** `.env` and not only a new one, so the documented path repairs itself. If you deploy by Helm, Terraform or your own compose, generate the three values first and set them on the `api`, `actions` and `realtime` services — `.env.example` documents each one and which services must share it.
-
-**What is measured, and what is not.** The prompt-injection guard was hardened and graded twice so the second number could contradict the first, and it does. On the corpus it was written against, detection went from 66.7% to **98.1%**. On 28 adversarial payloads authored after the guard was committed and never consulted while its patterns were written, it scores **7.1%, and that is the figure to plan against** — the hardening moved the corpus it was tuned on by 31 points and moved unseen payloads by a single payload. Both rates are published with the distinction stated and the held-out one leads, because a reader who stops after the first number leaves with the rate that describes the guard least. The held-out corpus deliberately carries no floor and no ratchet: a target on a held-out set is an instruction to tune against it. **No hosted provider has ever been exercised** — that still needs a funded key which does not exist, and it remains a separate, unmade claim. The npm and PyPI packages are built and packed on every tag and **none of the eight is uploaded**, because the first npm upload of a package needs a token no repository variable can replace; treat any install command for them as unavailable rather than broken.
+---
-**v12.0.0 highlights (September 28, 2026)**
-- **A stock install dispatched response actions to anyone who could reach the port.** The actions service skipped authentication entirely whenever its service token was empty and `AISOC_DEV_MODE` was set — and compose defaults that flag to `1` while nothing generated the token, so `isolate_host`, `disable_user`, `block_ip` and `run_script` were available unauthenticated on the profile `make up` starts. The realtime edge was the same shape twice over: connection tickets were verified against a constant committed to this repository, and the `/internal/*` guard read an unset token as authorized rather than as unconfigured.
-- **A `viewer` token could write to cases.** `cases:write` was withheld from the role deliberately and then enforced on none of the nine write routes in `cases.py`, so the read-only role could mutate a case and launch an investigation. A permission that exists, is documented, and is checked nowhere is indistinguishable from one that was never granted until somebody tries it.
-- **Hunting no longer starts from a hunt somebody already wrote.** An analyst gives a hypothesis and the agent fills a plan whose fields and operators are closed enums in the schema it is handed — 17 fields, 8 operators — which a compiler renders into SQL with every model-supplied value bound as a parameter. There is no string the model can produce that becomes SQL text, so injection is not filtered, it is unrepresentable. A lake that could not answer returns a reason and no `rows` key at all, so findings, no findings and could not check stay three outcomes rather than two.
-- **The hunt corpus went from 5 hunts to 68, and the grading covering it stopped being satisfiable by accident.** A hunt passed when it fired on its positive scenario and not on its negative, and the second half was free: a negative drawn from a different log source never fires on any hunt, so a corpus of those reports a perfect false-positive rate while grading only the hunt's aim. A negative must now differ from its positive in exactly one indicator field. That found **8 violations on its first run against a corpus the existing grading was passing**, three of them in hunts predating this work. `scripts/run_hunt_evals.py` also now exists: `hunts/README.md` had told contributors to run it before opening a PR for months, and it had never been written.
-- **Autonomy is earned from a measured track record rather than typed into a setting.** The tier, the per-action thresholds and the force-auto override were all numbers somebody entered, and nothing connected what the agent had actually got right to what it was allowed to do unattended. A grant now needs a minimum sample with enough of it closed as malicious, agreement whose denominator excludes abstentions so declining cannot inflate it, a capped abstention rate, malicious recall that counts an abstention as a miss, and a trailing slice scored separately so drift arriving gradually is not absorbed by a 30-day average. Promotion refuses an unmeasured rate; demotion ignores one, because a week with no malicious alert is not evidence the agent got worse.
-- **Deactivating a user ended their sessions and left their API keys working.** The key resolver looked up the owning user with `is_active == True` and, finding nothing, fell through to a generic `api_service` role instead of refusing — so every credential a deprovisioned principal had minted for themselves went on authenticating indefinitely with nothing in the request or the logs looking wrong. Anyone testing deprovisioning would have checked the session, which did stop.
-- **The gate written for this exact bug class had never seen a route.** It read `app.routes` and filtered for `APIRoute`, but on the FastAPI version these services pin, `include_router` appends a single `_IncludedRouter` instead — so the check compared zero pairs, found zero problems and reported success over an app serving 456 operations. The defect it should have caught: `DELETE /api/v1/autonomy-policy/grants` had been unreachable since it shipped, taken by a `DELETE /{action}` handler declared earlier in the same router, answering 204 to the caller while deleting from the wrong table and leaving the tenant holding the capability.
+## v12.0.0 (2026-09-28)
-The full inventory — 44 entries — lives under `[12.0.0]` in [`CHANGELOG.md`](CHANGELOG.md).
+The release where two services stopped serving unauthenticated. Full detail under `[12.0.0]` in [`CHANGELOG.md`](CHANGELOG.md).
---
diff --git a/VERSION b/VERSION
index 4044f9086..77903b35f 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-12.0.0
+12.1.0
diff --git a/apps/docs/static/data/scoreboard.json b/apps/docs/static/data/scoreboard.json
index 58b8873ce..b4c7138e6 100644
--- a/apps/docs/static/data/scoreboard.json
+++ b/apps/docs/static/data/scoreboard.json
@@ -3,9 +3,9 @@
"_readme": "Append-only public benchmark scoreboard. Each row is one eval run. Substrate rows (no LLM, deterministic) are marked substrate:true and MUST NOT be quoted as live agent performance. Wet-eval rows (substrate:false) are appended by the T5.5 weekly CI workflow once it lands. Newest row first.",
"rows": [
{
- "date": "2026-09-27",
- "agent_version": "v12.0.0",
- "commit_sha": "b7099f42",
+ "date": "2026-09-28",
+ "agent_version": "v12.1.0",
+ "commit_sha": "c55d3afa",
"substrate": true,
"eval_mode": "substrate-only",
"mitre_accuracy": 0.97,
diff --git a/apps/docs/static/llms-full.txt b/apps/docs/static/llms-full.txt
index 46dc2a1c0..cb28d3e0e 100644
--- a/apps/docs/static/llms-full.txt
+++ b/apps/docs/static/llms-full.txt
@@ -3,7 +3,7 @@
<!-- Generated by scripts/generate_llms_txt.py. Do not edit by hand; run the script and commit the result. -->
> AiSOC is an open-source, self-hostable AI Security Operations Center, MIT
-> licensed and maintained by the AiSOC community. Version 12.0.0.
+> licensed and maintained by the AiSOC community. Version 12.1.0.
> Repository: https://github.com/beenuar/AiSOC. Documentation: https://beenuar.github.io/AiSOC/.
This file exists so an automated reader can answer questions about AiSOC
diff --git a/apps/docs/static/llms.txt b/apps/docs/static/llms.txt
index 63984cf4a..b8805a01e 100644
--- a/apps/docs/static/llms.txt
+++ b/apps/docs/static/llms.txt
@@ -6,7 +6,7 @@
> security telemetry from 84 first-party connectors, normalises it to OCSF,
> correlates it into alerts, and runs LLM agents that triage and investigate them.
> Every agent decision is written to an investigation ledger, so the reasoning
-> behind a verdict can be replayed step by step. MIT licensed. Current version 12.0.0.
+> behind a verdict can be replayed step by step. MIT licensed. Current version 12.1.0.
AiSOC is built and maintained by the AiSOC community. It is a complete SOC
platform rather than a single tool: ingestion, detection, correlation, agent
diff --git a/apps/web/package.json b/apps/web/package.json
index 9432eadf5..667a5ab53 100644
--- a/apps/web/package.json
+++ b/apps/web/package.json
@@ -1,6 +1,6 @@
{
"name": "@aisoc/web",
- "version": "12.0.0",
+ "version": "12.1.0",
"private": true,
"description": "AiSOC SOC Console - Next.js 14 Frontend",
"license": "MIT",
diff --git a/infra/helm/aisoc/Chart.yaml b/infra/helm/aisoc/Chart.yaml
index fe0a86ec5..f8198e071 100644
--- a/infra/helm/aisoc/Chart.yaml
+++ b/infra/helm/aisoc/Chart.yaml
@@ -2,13 +2,13 @@ apiVersion: v2
name: aisoc
description: AiSOC - the open-source, AI-native Security Operations Center
type: application
-version: 5.6.0
+version: 5.7.0
# Used as the image tag wherever a `tag:` in values.yaml is left empty, so
# it has to be a tag the registry actually holds — it read "5.2.0", which
# exists for no image, so every pod of a default install would have gone to
# ImagePullBackOff. scripts/check_published_images.py resolves it the same
# way the templates do and fails when it names nothing.
-appVersion: "v12.0.0"
+appVersion: "v12.1.0"
keywords:
- security