Skip to content

Releases

Security Watch 1.2.0

Choose a tag to compare

@VODuda VODuda released this 19 Nov 17:00
a7ef557

News that no advisory has scored yet gets a level from what the articles say, and findings can show EPSS. Existing configurations and inventories work unchanged. To be published under the default latest tag; the GitHub Action is used as uses: ShieldaAI/security-watch@v1.2.0.

Added

  • News signals. For a news finding with no advisory severity (a zero-day without a CVE, a CVE that is reserved but not published, a vendor bulletin), the articles are read, without a model, for a CVSS score, a vendor's word rating, a report of exploitation, a denial of exploitation and public exploit code. A score or rating raises the finding's priority to at least its level; a report of exploitation raises it to critical, so with urgent_unknown (the default) the finding is sent at once as possibly critical. A statement counts only when it is tied to the finding's subject (its CVE or GHSA identifier, or the product or package it matched); negated and hedged statements and "zero-day" alone are not reports. Signals never lower, dismiss or mark a finding as known exploited, and stop being read once an advisory with a severity joins the finding. Alerts and digests have a News signals: line with the score, rating and subject only; explain shows the sentences they were read from. See news signals.
  • policy.news_signals (default true) turns news signals off.
  • EPSS, optional. sources.epss: true looks up FIRST's EPSS for the CVEs of each current finding (including the CVE aliases of a GHSA-keyed library finding) and shows the highest, with its CVE, in alerts, digests, explain, findings and the scan report (Assessment.epss, exported type EpssScore). Only CVE identifiers are sent to api.first.org, in batches of 50, at most ten requests a scan, each score refreshed once a day. It is shown only: it changes no priority or decision, and a new daily score creates no finding revision, reopens no review and asks no model for a new analysis. A failed lookup is a notice, and findings keep the last score. See EPSS.
  • Assessment.news_signals and the exported types NewsSignals and NewsStatement.
  • The configuration guide lists feeds that carry full articles with scores or exploitation reports (Rapid7, Tenable, The Register, Security Affairs, the Zero Day Initiative) and those that carry only headlines.

Changed

  • Louder for unscored news. A news finding that an article reports as exploited, or that an article scores at 9.0 or above, now alerts at once where 1.1 put it in the digest. Measured on one week of real articles against a 20-product inventory: five findings were raised and one more immediate alert was sent (an actively exploited ScreenConnect plugin). The first scan after the upgrade reassesses news findings still in the news window, so a raised one is sent once.

Not verified

  • Measured on 256 articles from one week of November 2026, labelled by the developer, not by an independent reviewer. The human-labelled pilot and a real Slack and email delivery remain open, as in 1.0.

Security Watch 1.1.1

Choose a tag to compare

@VODuda VODuda released this 19 Nov 15:33
e8025f9

Fixes for feeds that were read wrongly and for two checks that trusted a name. No configuration change. To be published under the default latest tag; the GitHub Action is used as uses: ShieldaAI/security-watch@v1.1.1.

Fixed

  • Feeds with an image in <media:content> were read without their text. The article in <description> was ignored whenever an item also had a <media:content> element, so every item from feeds such as Rapid7's blog and Dark Reading arrived with an empty description and could match only by its title. The first of content:encoded, content, description and summary that holds text is now used. An empty content:encoded also falls through to the description.
  • Feeds that escape their HTML with entities showed the tags as text. A description written as &lt;p&gt; (Tenable's blog, the seclists.org lists) arrived as the literal text <p>, <br> and similar, which reached matching, models and alerts. A field that carries HTML (RSS description and content:encoded, Atom content and summary with type="html") and holds no raw <, which escaped XML text never does, is now unescaped once as XML (the five predefined entities and character references only; DTDs are still refused) before it is read as HTML. CDATA and plain-text fields (titles, Atom text content) are read as before, so &lt;script&gt; in a title stays text. The first scan after the upgrade records a new revision of items from such feeds that are still in the news window.
  • A name that starts like a loopback address no longer counts as one. intake.listen: 127.example.com:8787 was accepted without TLS or allow_plain_http, because the check read the text of the host. Only localhost, an IPv4 address in 127.0.0.0/8 and the IPv6 loopback address count now.
  • sbom --project FILE refuses a symbolic link. A link named like a dependency file was followed to whatever it pointed at, unlike files found inside a project folder, which were already skipped. A project folder given as a link is still read.

1.1.0

Choose a tag to compare

@VODuda VODuda released this 16 Nov 16:33
f052b74

Start from your code instead of a hand-written inventory, and keep the inventory current from CI. Existing configurations and inventories work unchanged. To be published under the default latest tag. The GitHub Action is used as uses: ShieldaAI/security-watch@v1.1.0.

Added

  • security-watch sbom reads the lockfiles and manifests in a project and writes a CycloneDX 1.6 SBOM: npm (package-lock.json v1 to v3, yarn.lock 1 and 2+, pnpm-lock.yaml 5 to 9, package.json), Python (poetry.lock, uv.lock, Pipfile.lock, pdm.lock, requirements*.txt, pyproject.toml), Go (go.mod), Rust (Cargo.lock), PHP (composer.lock, composer.json), Ruby (Gemfile.lock, Gemfile), Java (pom.xml, gradle.lockfile), .NET (packages.lock.json, project files, Directory.Packages.props), Dart, Elixir, Swift and Haskell. It reads files only: no package manager, no symbolic links, no network. Each package records the files it was found in, and alerts show them (Affected item: npm express 4.21.2 - found in package-lock.json). --owner, --note, --environment, --exposure, --criticality and --system-id describe the system; in GitHub Actions and GitLab CI the repository address and commit are filled in. --production leaves out development-only dependencies. The command exits 1 when it finds no packages.
  • init --project PATH creates a monitor for one or more projects in one step: an SBOM per project in sboms/, an inventory that reads that folder, and OSV turned on when a project has packages outside npm.
  • SBOM folder. sbom_directory in the inventory makes every .json SBOM in a folder a system, read again at every scan. The SBOM's creation time is the system's updated_at, so an SBOM that stops arriving is reported as stale. A file that cannot be imported is reported and left out without stopping the other systems. Operating-system packages in container SBOMs (deb, rpm, apk and similar) are listed once as a notice instead of degrading every scan.
  • SPDX import. import-sbom, the system sbom field and the SBOM folder accept SPDX JSON 2.2 and 2.3, including GitHub's dependency-graph export, as well as CycloneDX. SBOMs from Trivy, Syft and cdxgen are accepted as they are.
  • SBOM receiver. An intake section makes start accept PUT /v1/sboms/SYSTEM_ID from pipelines, with bearer tokens limited to system IDs (lower case, as sbom writes them), TLS (or plain HTTP on loopback, or behind a TLS proxy with allow_plain_http), a 20 MiB limit, a slowdown after repeated failed tokens, and an answer a pipeline can act on: the upload is imported before it is written, so an unusable SBOM fails the job at once. security-watch intake runs the receiver alone. security-watch upload and sbom --upload URL send an SBOM, retrying network errors and busy answers; the GitHub Action takes upload-url and upload-token. An empty SBOM folder is accepted, so a monitor can start before the first upload.
  • GitHub Action (action.yml): uses: ShieldaAI/security-watch@v1.1.0 writes the SBOM and uploads it as an artifact, or checks and uploads an SBOM made by another tool (sbom-file).
  • Watch a project from its dependency files: supported files, CI recipes, ways to deliver SBOMs to the monitor, and other generators.
  • Library exports: generateSbom, importSbom, DEPENDENCY_FILES, startIntake, createIntakeHandler, uploadSbom, and the types SbomOptions, GeneratedSbom, DependencyFile, IntakeUpload, IntakeOptions, RunningIntake and UploadOptions. SbomImport.source.format may be SPDX, and SbomImport.subject describes the system an SBOM is about.

Changed

  • CycloneDX components with evidence.occurrences take the listed locations as their note.
  • The inventory's systems list may be empty when sbom_directory is set; an inventory with no system at all is still refused.
  • The README starts from init --project; the npm description and keywords say that lockfiles and SBOMs are read.

What's Changed

  • Release 1.1.0: SBOMs from lockfiles, SBOM folder, SPDX import, receiver and GitHub Action by @VODuda in #16

Full Changelog: v1.0.1...v1.1.0

Security Watch 1.0.1

Choose a tag to compare

@VODuda VODuda released this 15 Nov 22:40
1a0d943

Package metadata and documentation only; no change to behaviour or to the public API. To be published under the default latest tag.

  • The npm description now says what the tool does: "Self-hosted monitor that checks new security advisories and news against the libraries and products you run, and alerts you on Slack or by email." The keywords no longer suggest an npm-only, OSV-based tool (npm, osv and typescript replaced by security, sbom, email-alerts and nodejs).
  • The 1.0.0 notes below record its publication.
  • npm run release -- --dry-run was rehearsed on 2026-11-15 on commit 284ab8b (tag latest, 110 files, 577 kB packed, nothing uploaded); the commit that records it changes only this changelog.

Security Watch 1.0.0

Choose a tag to compare

@VODuda VODuda released this 15 Nov 22:23
43db188

First stable release, published under the default latest tag. Not yet published at the time of writing.

Upgrading from 0.1.0-alpha.0

  1. Use Node.js 24.10.0 or newer (the alpha accepted 24.0).
  2. Stop the old worker and take a backup: security-watch backup --output PATH.
  3. Install the new version, then run security-watch scan or security-watch start once. This migrates the database to schema 7. Read-only commands refuse to run until that has happened. There is no downgrade path other than restoring the backup with the old version. Findings that did not change keep their revision, their delivered alerts and their acknowledgements: the upgrade does not send them again.
  4. Run security-watch doctor. It now reports an SMTP address it cannot use, a verification model identical to the analysis model, and inventory entries that were left out (import_warnings) and npm names that may be a typing mistake (name_notices). It exits with status 2 whenever it prints ok: false.

Behaviour that changes for an existing installation:

  • AI dismissals are reviewed by default. News that a model classifies as general news or as not affecting you goes to a local review queue and the digest instead of being recorded silently. Set policy.ai_dismissals: record to keep the old behaviour.
  • Models receive less. Each model defaults to the minimal disclosure profile: package and product names and versions only.
  • models.verification must differ from models.analysis. A pair that reaches the same model at the same address is rejected at start-up, however each one is written (a spelled-out default endpoint, a trailing slash, different letter case, a compatible-endpoint profile pointing at the same service).
  • models.daily_token_limit defaults to 1,000,000 (was 100,000), and a request now reserves one token per three bytes instead of one per byte. The old defaults allowed about four requests a day.
  • Plain http:// endpoints (models and the external heartbeat) are accepted only for local or private hosts, even with allow_private_endpoint: true. An address is judged by its range; addresses that wrap a public IPv4 address (NAT64, 6to4) count as public for this rule, and HTTPS requests to them are refused. A name is accepted when it has no public form (ollama, ollama.ml-tools, *.internal, *.ts.net, *.test, *.home and similar), and the request is refused if the name resolves to a public address.
  • Inventory text is one line. System id, name, owner, environment, keywords, product names and every note are rejected when they contain a control character (tabs and line breaks included) or a Unicode line separator. These values are printed in notifications.
  • News is matched to packages by stricter rules. A package whose name is an English word (express, moment, request) no longer matches a story by the word alone; it needs to be named as a package ("the request package", "npm install send", "the npm package next"), a release, or to be written as a name. An existing news finding that rested on such a word becomes inactive at the first scan after the upgrade. Advisories, which are matched by structured package claims, are not affected. The word "module" no longer counts as package context.
  • A capitalised package name typed by hand is pointed out (npm, RubyGems, Hex, Pub). npm:Lodash is accepted, because some packages from before 2017 keep their capitals (MD5, CSSselect), but it matches only advisories for exactly that spelling. doctor lists such entries under name_notices and every scan report carries them in notices. A notice is not a fault: the scan is not degraded, no health alert is sent, and the heartbeat is not withheld. Names imported from a lockfile or SBOM, and JSONStream, Base64, JSONPath and JSONSelect, are taken as published.
  • cve.now reads one week past its checkpoint on every scan (about 45 listing requests at the current rate of some 600 records a day, under a minute) to find records the service ingested after their publication date.
  • The first cve.now scan looks back one day (was seven). A week meant about 4,500 detail requests and 80 to 90 minutes before the first result. Exact npm package matches still cover all of history.
  • Recipient domains are compared in lower case. Email to Ops@Example.ORG was delivered but recorded as uncertain.

Added

  • Notes that say where to look. Systems, libraries and products accept a note; library list lines accept # note. A notification about a system carries a Where to look: line with the system name, owner and notes, and one about a library or product also an Affected item: line. Notes are never sent to sources or models.
  • CycloneDX 1.4 to 1.6 import, per-source coverage receipts, inventory drift, declared artifact and remediation evidence.
  • CVSS v4 base-vector scoring for OSV and GitHub records.
  • Local review queue for AI dismissals (reviews, review).
  • Offline rehearsal (rehearse) and pilot measurement commands (pilot-report, pilot-cases, pilot-score, pilot-recall, pilot-response-template).
  • acknowledge --revision N refuses to acknowledge a revision newer than the one that was read.
  • Library exports: safeHttp, acquireLock, notificationText, sendEmail, smtpProblem, attachLocation, SCHEMA_VERSION, and the types HealthReport, ExternalHeartbeat, MaintenanceRun, EmailSender, EmailMessage, EmailReceipt, DeliveryOptions, FindingLocation. deliverPending accepts an options object.

Fixed: alerts that were wrong or lost

  • A validated AI "affected" verdict on a news finding could be recorded silently while an uncertain one alerted. It is now never quieter than the uncertain route.
  • A news finding matched by keyword or package name disappeared when an advisory record arrived for the same CVE without a package claim this application can use (no claims at all, an ecosystem it does not know such as GitHub Actions or Debian, or a name it cannot normalise). It now stays and takes on the advisory's severity and exploitation status.
  • While a source was incomplete, an excluded finding flipped between "not affected" and "unknown" on every scan and sent a new alert each time. Only the advisory source that supplied the excluding version range counts: a linked news article that could not be read no longer turns an exclusion into an urgent "unknown".
  • After the monitor had been stopped for longer than two poll intervals, the first scan briefly stored excluded findings as "unknown" and queued urgent alerts for them before refreshing the evidence. Stored evidence is now judged as of the last collection until it has been refreshed.
  • With a model enabled, editing a note (or the owner) created new revisions, a new review beside a dismissed one, and a second digest or immediate alert. Notes, the owner and the Where to look text are now left out of everything that decides revisions, reviews, analysis reuse and alerts.
  • Upgrading from 0.1.0-alpha.0 gave every active finding a new revision and a new alert and dropped acknowledgements, because the stored fingerprint gained fields. A finding unchanged in what the earlier release compared keeps its revision.
  • CISA KEV evidence was not attached to news about a CVE that had no advisory record yet, so the known-exploitation override did not fire.
  • npm ranges using grammar that advisory sources do not use (^, ~, hyphen ranges, > 1.2) could produce a false "not affected". They now give "unknown".

Fixed: sources that stalled

  • The set of stored cve.now records re-read at every scan grew without bound and eventually failed the source. Records without a package match now take turns, 200 per scan.
  • A backlog larger than one scan reads (cve.now, GitHub) was never drained. Collection now resumes from the last record processed.
  • One package name that could not be normalised discarded the whole advisory and held back the checkpoint.
  • A stored OSV record that no longer exists stopped the refresh of the records after it.
  • Feed entries that appeared after their own timestamp were dropped silently.
  • cve.now changed its response formats: the package lookup is now a JSON Feed and the listing pages with next_cursor. Both the earlier and the current formats are read; before, the package lookup failed and only the first listing page was read.
  • A cve.now record stamped later than the scan's start time (for example one updated while a long first scan was still running) made the whole source incomplete. Such a record is now collected with its earlier timestamp; only a record with no usable time is refused. A listing record with a publication date ahead of the clock is collected and never ends paging.
  • cve.now records ingested after their publication date were never collected: the listing is ordered by publication date, so they appear below the checkpoint. The listing is now read one week past the checkpoint and a record counts as new when it was first seen inside the window. A record ingested more than a week late is found only through an exact npm match or a news citation.
  • One cve.now record could stop the source for good: a stored record the service no longer serves (HTTP 404) failed every scan, and a record that cannot be parsed held back the resume point of a backlog. A new record that cannot be read is asked for again until its requests have failed three times and, since the first failure, one day has passed when the service answers with a refusal or with content that cannot be used, or three days when the request itself fails. Then it is reported and given up on. Only the scan of its first failure holds the checkpoint, and not even that one when it listed more than 200 new records; after that the record is remembered by its identifier and the checkpoint moves on, so no scan has to read the same listing again because of it. A service that allows only so many request...
Read more

Security Watch 0.1.0-alpha.0

Pre-release

Choose a tag to compare

@VODuda VODuda released this 24 Oct 21:21
abde870

Security Watch's first alpha monitors security advisories and news for a configured inventory, with optional model analysis and Slack or SMTP alerts.

Install with Node.js 24 or newer:

npm install --global @shieldaai/security-watch@0.1.0-alpha.0
  • Import libraries by name or exact version across 12 ecosystems, including PHP, Python, npm, Go, Java and Rust. Missing versions remain uncertain.
  • Configure the scan interval separately from the news window; setup defaults to a scan every 30 minutes covering the last 40 minutes.
  • Use 22 provider profiles across six API protocols, with bounded requests, evidence checks and persistent budgets.
  • Keep processing valid source records, findings and queued alerts after isolated failures. Delivery retries distinguish explicit rejection from uncertain receipt.
  • Check worker health, configure an external heartbeat, preview retention and verify database backups before rotation.

The release commit is abde870beb5058d8e82a071df11afb0ba29aecda. CI and CodeQL passed, as did all 639 tests, formatting, strict compilation and clean package consumer checks. The public npm tarball matches the inspected release artifact byte for byte; its installed CLI, setup, library normalization, offline matching, types and runtime exports were checked separately.

This is an alpha. Provider protocol tests use simulated responses; they do not establish live compatibility for every hosted model. Prior live OpenAI checks passed schema and evidence validation; the latest Anthropic evaluation was blocked by invalid credentials. Slack and SMTP were checked with simulated transports; no real notification was sent. See the release review for source checks, evaluation results and operational limits.