watchdep
Sign inSource
Docs / Concepts / 10

AI review

Section
Concepts
Page
10 of 16
Reading time
3 min

The ranking rules say how bad an advisory is in general. They cannot read what the advisory says: the conditions it states (“only when the Socket Appender uses TLS”), a fix that was incomplete (Log4j 2.15.0), a later advisory on the same component that moves the version to upgrade to, or text that contradicts its own data. An AI review reads exactly that, for each finding, with two models: Claude Opus 5.5 and Gemini 4 Carbon.

What a review can and cannot do

A review only adds attention.

  • It may explain what the finding might affect in your application, which of the advisory’s conditions your inventory suggests are met, what the rules may have missed, and the fix that actually ends it.
  • It may suggest a higher level. It never lowers one: a lower suggestion is held at the rules’ level, and the page says so.
  • It never changes a level or a state by itself, and never closes, accepts, acknowledges or marks anything not affected. A person decides.
  • Every claim either quotes the advisory or says “general knowledge”. A quote that is not word for word in the advisory is shown as unverified.
  • A missing or failed review is never “nothing to worry about”. It says “AI review unavailable” and why, and alerts and ranking carry on exactly as without it.

watchdep sees your dependency inventory, never your code, so most conditions come back as “unknown”, with what to check. watchdep check --assist reads the code where it is.

When findings are reviewed

  • Automatically when a finding becomes new, reopens or escalates at severe or high (the owner can choose other levels).
  • On request: “Ask for a review” on any open finding.
  • Again only when something the review was given changed: the advisory or a related one, CISA KEV or EPSS, your inventory, a code check, or the model. “Review again” forces one.

Reading a review

The finding page shows both reviews side by side: each one’s suggestion (“Agrees with the rules: high”, or “Suggests severe (rules: high)”), its confidence, the impact here, the conditions (likely, unlikely or unknown), what the rules may have missed, and the fix, each claim with its quote.

A finding is worth a look (a blue eye in the lists, and a filter) when either review suggests a higher level, or the two disagree.

Raising a level

Beside a higher suggestion, members, admins and owners see Raise to severe (or whichever level). Raising is your decision: it needs a reason (prefilled from the review), it is logged under your name, it alerts as an escalation, and the rules never take the level back under it. Remove the raise returns the rules’ level.

What is sent

Only while AI review is on, for a finding being reviewed: the advisory’s text and related advisories’ ids and fixed versions; the package, its version, paths and scope; the level and why, KEV and EPSS; the part’s other packages (names and versions), runtimes and databases; code check verdicts, paths and line numbers (and snippets only if your owner switched them on). Names are replaced with “application A, part 1”. No source code is sent. The full list is on the Trust page.

Settings, AI review

The workspace’s owner can:

  • turn it off, which stops all sending at once;
  • choose who reviews: one or two of Anthropic, Google and OpenAI;
  • choose which levels are reviewed automatically;
  • add your own key for any of the three and choose its model. Reviews on your own keys run on your account and do not count against your plan’s monthly reviews. A key is encrypted, shown afterwards only by its last four characters, and sent only to its own provider. If a provider refuses it, you see why, and reviews use the service’s key while your plan allows.

Monthly reviews by plan

Free Team Business
AI reviews a month on the service’s keys 50 1,000 5,000

Two models on one finding count as one review. Billing shows this month’s use; the owner is emailed at 80 and 100 percent. When they are used, new findings show “AI review unavailable: this month’s AI reviews are used” until the 1st, and nothing else changes.