Products / Software / A440

Adversarial intelligence

A440

A440 investigates how authorized agents can violate their intended boundaries, demonstrates the effects in controlled environments, and verifies whether a proposed fix removes them while preserving useful work.

Request a demo What it does today

Teams ready to scope one can request a pilot.

Form
Software. A private local workspace.
Applies to
AI agents that make real tool decisions, and the controls around them
Availability
Version 0.3. Demos on request.
Name
A440 is the reference note an orchestra tunes to. Syntony's A440 brings an agent's claims, controls and evidence into agreement.
A440 home view, Find the dissonance: the composer at the center of the workspace with three starting actions below it, for probing an agent's boundaries, making sense of the evidence and designing a stronger defense.
A440 v0.3 / home view, desktopPrivate local workspace
A440 home view on a phone: the Find the dissonance prompt, the composer and the three starting actions stacked vertically.
Mobile
Version 0.3Current build
127Automated tests passing
26 / 123Research base: product families and primary sources reviewed

As of 2 November 2026.

01 / What it does today

Find the dissonance. Attack an agent, verify the fix, export the evidence.

A440 is a conversational workspace where a security team can probe an agent's boundaries, make sense of the evidence and design a stronger defense.

Under the conversation, it generates adversarial inputs with a live model, runs them against an agent making real tool decisions, checks the state changes that actually occurred, replays the exact attack against strengthened controls, and exports the evidence as a report, a JSON record and a regression suite.

It runs as a private local workspace.

  1. 01
    Generate. Adversarial inputs from a live model, shaped to the agent's tools and permissions.
  2. 02
    Run. Inputs go against an agent making real tool decisions in a controlled environment.
  3. 03
    Check. The state changes that actually occurred, not the agent's account of them.
  4. 04
    Replay. The exact attack against strengthened controls. Useful work must still pass.
  5. 05
    Export. A report, a JSON record and a regression suite your team keeps running.

02 / More views

A completed campaign. A verified control. An exported record.

Screens to follow from the v0.3 package.

COMPLETED CAMPAIGN
SCREENSHOT TO FOLLOW
Campaign viewTODO
CONTROL VERIFICATION
SCREENSHOT TO FOLLOW
Replay against a fixTODO
EVIDENCE EXPORT
SCREENSHOT TO FOLLOW
Report, JSON record, regression suiteTODO

03 / Scope and limits

What A440 is, stated plainly.

  • A440 works with a live model to generate adversarial inputs. It does not use Syntony-trained model weights.
  • It runs as a private local workspace for one team at a time. It is not offered as a hosted multi-tenant service.
  • Targets run inside controlled environments set up with your team. External target adapters and disposable VM execution are not part of the current build.
  • It finds the failures it is pointed at and verifies the fixes you propose. No tool finds every threat.
  • Research and engineering design brief available on request.

See A440 against an agent like yours.

One agent, one boundary, one proposed fix, one exported record. Bring the agent you are most worried about.

Request a demoBook a call