Stalwart AI Assurance
AI agents that do real work through your tools. They pick the right tool from what your users actually say, never write unless your rules allow it, and need no model call to choose.
Choosing a tool
No model call
Writes
Only when your rules allow
Guardrail bugs
Found, with the fix
Access
One API key
Everything an agent needs to act safely in your domain
Five products, one API key: choose the right tool from what your users say, find the bugs in your guardrails before anyone else does, write rules in plain English, and check your models behave as required.
Guardrail Bug Hunter
Find every way past your AI guardrails before an attacker or an accident does.
Guardrail Verification
A yes-or-no answer, with evidence, to “can this agent ever do X without Y?”
Guardrail Writing
Write a rule in plain English and get back a policy that has already been checked.
MCP Tool Selector
Predictable, explainable tool choice for AI agents.
Random Forest Verification
Show that your random-forest model behaves as required, or get the exact case where it does not.
Bot Builder
Declare a domain agent that acts with all of the above built in.
We scanned the public guardrails that a whole industry copies
We pointed the Guardrail Bug Hunter at the public rule files of a leading open-source AI guardrail framework: the framework's own library of ready-made safety rails, the sample assistants its vendor publishes, and partner and community projects. 185 files from 12 repositories, scanned offline with no AI model in the loop.
Two files held real vulnerabilities, and each was reproduced in the framework's own runtime. Both are fixable by the Bug Hunter: it writes a corrective rule for each, and checks that the rule closes the hole without blocking anything else.
These are professional reference files that many teams copy, so every bug travels with every copy. Red-teaming would have had to guess the right conversation. The Bug Hunter found each one, with the conversation that triggers it.
185
guardrail files from 12 public repositories
2
vulnerable files, each reproduced in the framework's own runtime
32
tool calls an attacker could trigger just by steering the conversation
8.5 s
to scan the framework's 115 toolkit files
Hallucinations reported as jailbreaks
A rail that detects hallucinations raises them as jailbreaks. A team that blocks jailbreaks blocks every hallucination too, and a team that handles hallucinations never sees this rail's.
Three defects in a food-ordering assistant
Clearing the cart never runs, swapping an item checks the wrong value, and replies about the customer's order are left to unchecked AI-generated text.
Tool calls anyone can trigger
32 tool calls in 28 files run on the user's apparent intent alone, so anyone who can steer the conversation can trigger them. Two of them place food orders with no check of who asked or whether they confirmed.
Figures from our research paper, for a scan run on 3 October 2026.
How an agent is kept to its rules
Low cost, relevant and safe by design, not by luck
A model left to run an agent pays for every decision, guesses at your domain, and writes whatever it decides. Stalwart AI Assurance puts your rules and your users' words in charge, and calls a model only where you choose.
| WHAT MATTERS | A MODEL ON ITS OWN | STALWART AI ASSURANCE |
|---|---|---|
| Choosing a tool | A model call every turn, and a different answer some days | No model call; the same request picks the same tool |
| Writing through your tools | Whatever the model decides | Only when your rules allow it; otherwise it asks or refuses |
| Your domain's words | Prompt tweaking, again | Learned from your users' feedback, versioned |
| Bugs in your guardrails | Found when someone happens to try them | Hunted down, each with its conversation and the fix |
| Paying for a model | On every decision | Only where your setting sends a turn to one |
Put an agent to work, safely
Read the API reference and try a call, or register your interest and our team will be in touch.
A finance case study, step by step
Our technical demo runs the engine on a regulated-finance policy set: a trade-booking approval gate, a below-threshold exemption that opens a bypass, and a consumer-duty advice rule. Pick a case to see the policy, the verdict and the evidence.
Algebraic Guardrail Verification in Action
define user request_trade_booking
"book this contract note"
define flow trade_booking_gate
user request_trade_booking
if $four_eyes_approved
bot book_trade
else
bot refuse_unapproved_bookingMATHEMATICALLY SAFE (UNSAT)
The algebraic solver derives 0 = 1 in 5 steps. Over all possible 2^N runtime contexts, the action book_trade is provably unreachable unless four_eyes_approved is positively asserted.
Algebraic verification benchmarks
Regulatory policies, tool dispatch gates and state lookbacks, decided by deterministic finite-field elimination.
FCA Consumer Duty Advice Guardrail (410 Variables)
410-variable synthesized GF(2) circuit under FCA COBS/Duty obligations. Proved that advising permissions cannot be bypassed via targeted-support advice flows.
Middle-Office Trade Booking Dispatch Gate
Discovered an OR-accumulation hole in a sub-threshold booking exemption. Generated a minimally disruptive policy patch certified to block the vulnerability.
Multi-Step Velocity Chain State Lookback ($prev.v)
Stateful lookback verification enforcing dynamic transaction frequency caps across conversational turn state variables.