See how AI risk moves through your system, and where it can be stopped.
A customer-service agent can retrieve documents and change accounts. As its permissions and usage grow, so can the ways it fails. Explore what changes when an approval boundary is added before privileged actions.
In development. We’re speaking with prospective pilot partners about their needs and potential fit.
Under the model
- Feedback loops found automatically
- –
- Stocks tracked
- –
- Intervention levels tested
- –
- Months simulated
- 48
What if account changes required approval?
The baseline represents an agent whose scope can expand while testing and control coverage struggle to keep up.
One intervention: add an approval step before a privileged account action. Compare the modeled behavior while keeping the other assumptions fixed.
Loading the illustrative comparison…
These are precomputed model scenarios. They suggest a hypothesis to test; they do not measure incident probability or predict results for your organization.
— Baseline— Selected scenario
Which account actions need an enforceable approval boundary? Test bypass attempts and legitimate requests before relying on it, and retest when tools or permissions change.
The workflow we’re building.
- Evaluate
Our red team finds the attack paths.
Syntony’s human-led red team and Trusted Red Team Network test your system and record each attack path with its evidence.
Product direction: receive findings as structured evidence for a system map.
- Map
Connect findings to the system.
The aim is to connect findings to the architecture, releases, controls and people around them, so a failure can be understood in context.
Product direction: model feedback loops and explore their contribution to system behavior.
- Govern
Connect evidence to a response.
We’re exploring how system maps can inform candidate interventions and governance records for the people accountable for the response.
Product direction: connect with GovTune AI, also in development, for human review.
- Assure
Revisit evidence as the system changes.
Model swaps, prompt edits, new tools and new corpora can invalidate earlier test results. We’re building toward a workflow that keeps these changes connected to the evidence.
Product direction: identify evidence to review and tests to run again.
Inspect the baseline system map, feedback loops and model behavior
Press play to watch control pass between loops.
Drag sideways to see the whole map.
What happens when no one intervenes
Illustrative system, 48 months.
Which loops are in control
Share of the system’s behavior explained by each family of feedback loops, month by month.
From an illustrative model to a useful product.
Built from evidence
Our direction is to build maps around an organization’s architecture and evaluation findings, with assumptions that can be inspected and tested.
Kept current
We’re exploring how model, prompt, tool and data changes should prompt updates to the map and the evidence it depends on.
Connected to decisions
The goal is to connect modeling results to owners, controls, decisions and re-test plans that teams can use.
Join the Resonance waitlist.
Resonance is in development. Leave your email for product updates and news about availability. Joining does not provide immediate access or promise a launch date.
Interested in Resonance for your team? Discuss your system. Our products are in development. We’re speaking with prospective pilot partners about their needs and potential fit. A proposed pilot would assess whether a map of one agreed workflow leads to useful tests and interventions. You can also explore Syntony’s red-team services.
This preview uses an illustrative customer service agent with retrieval and account tools. It contains no client data, incidents or measured probabilities. Resonance’s model structure, parameters and calibration methods are proprietary to Syntony; see our terms. The loop analysis builds on Meadows, Leverage Points (1999).