Finding deduplication assist
- Appears
- AI review, proposing identity merges the rules missed, which a person accepts or rejects
- Grounded in
- Identities on the same hosts from different scanners, with their titles and ids
PYXA AI
AI drafts the work. Your people decide it.
Fifteen AI tasks sit inside the screens where vulnerability work already happens. They draft fix instructions from your own past fixes, propose plans for findings nobody owns, group a campaign by the one change that fixes it, and turn a sentence into a triage rule. Each runs on the model you choose, local by default, and each stops at a proposal a person accepts. Remediation guidance is in every edition; the other tasks come with Professional and Enterprise.
Where the hours come back
The writing nobody has time for, drafted.
Most of a remediation cycle is not the patch. It is writing up the fix for each team, grouping hosts that share a cause, deciding which plan a new vulnerability belongs in, and documenting what worked. PYXA's AI drafts each of those where the work happens, from data your instance already holds.
Every draft below was written by qwen2.5:7b on Ollama, on the same network as PYXA, against the demo estate. Nothing left the environment.
Remediation guidance
The fix, written up for the team that applies it.
Asked to draft the Struts emergency, the model proposed upgrading to 6.4.1 or later on the 25 payments hosts still open, Windows first, following the team's recorded procedure FIX-001 in a change window. It kept PYXA's rule that the finding closes only when a verifying scan no longer reports it.
The plan's owner accepts the wording or rejects it. Accepting changes the title and description, nothing else: no status, no priority.
Action plan proposals
Plans for the work nobody owns.
1,796 open findings sat in no plan. The model proposed six plans to cover them; PYXA ordered them by risk, then blast radius, then compliance, then quick wins, and set nothing until a person decided.
A reviewer rejected three. Cisco IOS and ASA do not share a patch, and neither do MongoDB and PHP. The rejections stay on screen with their reasons, so anyone can see where the model was overruled.
Plain-language search
Say what you are looking for.
Typed "everything failing verification", the model mapped the sentence to the Failed verification view and said so before anything changed. A person clicks Apply. The model never filters a list on its own, and never answers from its own memory: it only picks among the filters the page already has.
Root-cause clustering, knowledge drafting, rule drafting from a sentence and AI triage work the same way, each as a proposal a person accepts. They draft best with a larger model, and Enterprise routes each task to the model that suits it.
All fifteen tasks
Every step of the workflow, one task at a time.
Each task is listed as the AI page lists it: where it appears, and what it is grounded in. Each can be routed to its own provider or switched off on its own.
01
Ingest
Pentest report reading
- Appears
- Penetration tests, proposing the findings a report describes, each with the sentence it rests on
- Grounded in
- The text of the report uploaded to that engagement, and nothing else
02
Prioritise
AI triage
- Appears
- AI review, proposing the plan an unplanned vulnerability belongs in, which a person accepts
- Grounded in
- Unplanned vulnerabilities with their band, fix, host OS and team, and the open plans
Filter translation
- Appears
- The query field above the filters on Evolution, Findings and Assets
- Grounded in
- The filter vocabulary this instance actually has
Attack path narration
- Appears
- Vulnerability profile, under a BloodHound path
- Grounded in
- The path's own nodes and edges
03
Plan and assign
Action plan proposals
- Appears
- Action Plans, proposing plans for open findings no plan covers
- Grounded in
- Open findings no plan covers, with host counts, priority and recommended fix
Root-cause clustering
- Appears
- Action plan detail, proposing groups over the plan's open vulnerabilities
- Grounded in
- The asset and finding graph this instance already holds
Auto-triage rule drafting
- Appears
- Auto-triage rules, turning a sentence into a draft rule a person reviews and saves
- Grounded in
- The scope vocabulary this instance holds: nodes, attributes, owners, OS names, plans, people
Assignment rule proposal
- Appears
- Asset assignment, turning a sentence into a draft rule
- Grounded in
- A sample of hostnames, tags and OS strings from your estate
Fix effort estimate
- Appears
- Vulnerability page, beside the recommended fix
- Grounded in
- The fix's own procedure text and the attempts recorded against it here
04
Fix and prove
Remediation guidance
- Appears
- Finding detail, below the host ledger, and a plan's What to do, as instructions its owner accepts
- Grounded in
- Your own recorded procedures and fixes, the vendor's advice, and past fix notes
Verification failure diagnosis
- Appears
- Verification failed, beside the claimed and observed columns
- Grounded in
- This failure and up to three past failures that look like it
05
Report and learn
Executive summary drafting
- Appears
- Command Center and the CISO one-page, as an editable opening
- Grounded in
- This period's own figures, the ones on the page
Knowledge entry drafting
- Appears
- Knowledge, offered when a campaign reaches verified
- Grounded in
- That campaign's own timeline, deploys and rescans
Exception justification review
- Appears
- Nowhere. Switched off deliberately
- Why
- "A formal control should not be pre-judged by a model. Accepting risk is a decision with a named owner and an expiry."
The data boundary
A routing table, not a chat window.
Bring your own provider in any edition: a local model (Ollama or any OpenAI-compatible endpoint), a LiteLLM proxy, OpenAI or Anthropic. Enterprise routes each task on its own, so sensitive tasks can stay local while others use a larger model. The table says, task by task, exactly what would leave your environment. On a local model the answer is nothing.
- Remediation guidanceNo hostnames or addresses
- Verification failure diagnosisHostnames and addresses redacted
- Root-cause clusteringNothing, local
- Executive summary draftingNothing, local
Every prompt is kept in an audit log an administrator can read. Turning every provider off leaves a working product: no feature in PYXA requires a model.
Where AI stops
AI proposes. A person or a connector decides.
It cannot close a finding.
Only a verifying scanner observing the vulnerability gone can do that. A model's opinion that a fix worked is a claim, like a patch report.
It cannot approve an exception.
Risk acceptance is a decision with a named owner and an expiry, approved by someone other than the person who asked. Exception review is switched off by design.
It cannot change a priority band.
A band comes from the score's named contributions. AI can propose which plan a vulnerability belongs in; it cannot make it more or less urgent.