Phoenix / Products / Talon
Advanced offensive validation · operator-driven · authorized

Phoenix Talon

Ask Core to prove it. Then authorize it.

Can we go deeper than a scan? Talon is Crucible's deep, operator-driven mode. From Phoenix Core's chat, an authorized operator asks for a full engagement in plain language. Talon scopes and plans it — but nothing active runs until the operator clears an explicit authorization gate. It then drives a real recon → vulnerability → proof-of-exploit → post → report chain against your own estate, correlates every finding against current exploited-in-the-wild intelligence, and hands back evidence: which weaknesses are genuinely exploitable, and exactly how to close them.

target & evidence data sent outside your perimeter0
Launch from Core — gated by authorization

You ask in chat. It plans. It does not arm until you authorize.

The capability is deliberately two-stage: planning is free, acting is gated. A request alone never touches a live target.

  operator asks in Core chat    "run a full engagement against <our scoped target>"
                 │
                 ▼
  scope & plan  (read-only)    targets confirmed · phases laid out · nothing active yet
                 │
                 ▼
  ┌────────────  AUTHORIZATION GATE  ────────────┐
  │  operator clears an explicit, scoped engagement     │
  │  authorization. No gate cleared → no active step.   │
  └─────────────────────┬───────────────────────────────┘
                 ▼
  active validation             recon → vuln → proof-of-exploit → post
                 │
                 ▼
  evidence & report           what actually breaks · prioritized remediation
By design, not by policy. Talon will not arm an active technique until the authorization gate is cleared. Engagements are scoped to assets you own or are contracted to test, the read-only and active phases are separated, and every action is written to an audit trail. Launching is admin-only and takes a typed confirmation phrase. Exploitation sits behind a separate server-side switch that is off by default. A scope blocklist is enforced, and only one engagement runs at a time. The gate is a consent control the operator holds — never a bypass, and never implicit.
The engagement chain

From "you might be vulnerable" to "here is what actually breaks."

Recon

Map the real attack surface

Service and port discovery across the scoped targets, written to one source of truth for the run.

Vulnerability

Reason about what matters

Findings are matched against a local CVE corpus and analysed by the model — version-aware, not a blind signature dump.

Proof-of-exploit

Prove it, don't assume it

Where authorized, Talon safely demonstrates that a weakness is genuinely exploitable — the difference between "CVSS 9.8" and "this one actually lets someone in." Execution is guarded and staged.

Post & report

Evidence a human can act on

Curated post-exploitation recipes, then a findings summary, threat-intel context and a narrative — artefacts, not a 400-page export.

Model-agnostic & extensible

Bring your own model. Load the skills the engagement needs.

Any Ollama-compatible model endpoint

Talon works with any model served over an Ollama-compatible endpoint, local by default. It has run engagements on several different open-weights models and isn't tied to one vendor.

Skill-driven runs

An engagement follows a written skill that lays out the phases, the tools and the checks. The model works through it step by step and doesn't improvise a plan.

Proven tooling underneath

Standard, battle-tested frameworks drive the mechanical work — Metasploit RPC, curated scanners and post-ex kits — orchestrated, not reinvented. They include Active Directory (impacket), web discovery and SQL injection (ffuf, sqlmap), and Nuclei templates.

Intelligence, not guesswork

Exploited-in-the-wild first.

A finding with a known in-the-wild technique outranks a higher theoretical score that nobody is using.

Offline CVE corpus

~159k entries, on your box

CVSS and descriptions come from a local CVE database and a local mirror of the official CVE List, so there are no rate-limited calls out to a public API mid-engagement.

KEV-tagged prioritization

Known-exploited, surfaced first

Findings carry exploited-in-the-wild tags, so remediation effort goes where real attackers are already operating.

Built for authorized use, by construction

An offensive-validation tool for defenders — not an attack platform.

▪

The authorization gate is mandatory

Active techniques do not arm until the operator clears an explicit, scoped engagement authorization. Planning and recon can be read-only; exploitation cannot be implicit.

▪

Scoped to what you own

Engagements run against assets you own or are contracted to test. Scope is set up front and enforced, not assumed.

▪

Everything is logged

Each run keeps one source of truth — phases, attempts, sessions, findings — so an engagement can be reviewed, reproduced and defended after the fact.

▪

Nothing leaves the perimeter

Talon runs on your own infrastructure. Target data and evidence stay inside your network; connecting an external model is an explicit operator choice.

What's live today

Shipped, and what's rolling out — stated plainly.

CapabilityStatus
One CLI: recon, scan, vuln analysis, reportingLive
Offline CVE corpus + KEV taggingLive
Metasploit RPC + curated post-ex recipesLive
Launch from Core chat, behind the authorization gateLive
Model-agnostic (Ollama-compatible), skill-driven runsLive
AD, web & SQL-injection toolkits (impacket, ffuf, sqlmap) · Nuclei templatesLive
Identity-pinned reports (verified hostname and OS)Live
Targeting from the host's real installed-package inventoryLive
Guarded proof-of-exploit for non-Metasploit CVEs (sandboxed, off by default)Live
Fully autonomous multi-host campaignsRoadmap
Get started

Prove what actually breaks — on your estate, under your authorization.