One poisoned update. Developers' AI tools told to hunt for secrets.
4 hours
malicious Nx versions were live on npm
For teams putting AI agents to work
Every agent action is checked against your rules before it runs. Risky ones wait for your approval. Your auditors get proof, and your data never leaves your company.
The risk
AI agents now move your money, change your records and use your tools. Most companies only find out what happened afterwards, from the logs.
4 hours
malicious Nx versions were live on npm
~8 hours
an unauthorized Cline release stayed live
CA$812.02
ordered by a tribunal over its website chatbot's answer
Source: Civil Resolution Tribunal of BC, Feb 2024 (opens in a new tab)
With TraceStone, nothing runs before it is checked.
Why now
Regulators now expect you to control and record what your AI does. In Singapore, the MAS Guidelines on AI Risk Management set clear dates for financial institutions.
Key AI dates, Oct 2026 to Dec 2028
Sources: MAS Guidelines on AI Risk Management (Oct 2026) · MAS SAFR paper v1.1 · EU AI Act
How it works
TraceStone runs inside your company, between your AI agents and the tools they use.
Your rules
Your signed records
READ revenue allowed DELETE customer_records blocked PAY S$48,000 waiting you approved PAY S$••••• → ••••• Rules held No sensitive data
reaches the auditor
Runtime control
Set your rules once. Every agent action meets them before it touches your systems.
Example: an AI agent tries three actions. DELETE customer_records breaks the rule "no deletion", is blocked and becomes record #005. READ revenue is allowed and becomes record #006. PAY S$48,000 to K. Ong is over S$10,000, so it waits for the CFO; the CFO approves, the payment is sent and becomes record #007.
Tamper-proof records
Every agent action is recorded and sealed. If anyone removes or changes a record, you and your auditor see it at once.
An admin deletes one record
Examplee9514218
hash 7f7ac212 7f7ac212
hash 42866251 42866251
hash 39660d73 39660d73
hash eb124917 eb124917 Your audit
Your auditor asks for proof that every rule was followed. With TraceStone, you hand over proof, not your customers' names and amounts.
Learn more about private audit
Payment
SENTto K. Ong
BLOCKEDby the security tool
Waiting for a person
BLOCKEDyour CFO said no
Record
14:02:11 PAY 48000 okPlain log, easy to change PAY 48000 · blockedLog in the tool's console Your auditor asks: did every payment over S$10,000 have approval?
What your auditor gets
Full logs. Every name and amount is visible.
Yes. Every payment over S$10,000 had approval. Zero-knowledge proof
+ 1,237 more payments
Logs ask your auditor to trust you. Proof lets them check.
Why TraceStone
Cloud platforms protect agents inside their own platform. Security tools cover many platforms, but your auditor still reads raw logs. TraceStone covers all your agents and gives your auditor proof.
| Criterion | TraceStone | CrowdStrike Falcon Guardian | AWS AgentCore Policy | ServiceNow AI Control Tower |
|---|---|---|---|---|
| What your auditor sees | Proof that the rules held, not the customer data | Raw records, with redaction options | Logs, protected by access control | Records inside the platform |
| Where your auditor checks | Anywhere, offline, with a standalone verifier | In CrowdStrike's console | In the customer's AWS account | Inside the ServiceNow instance |
| Tamper-evident records | Each record is signed and linked to the one before |
Not stated
(audit trail only) | Only for activity logs, not decision logs |
Not stated
(audit records only) |
| Task-aware rules | Rules derived from each user task, verified before use |
Not stated
|
Not stated
(fixed rules, checked with automated reasoning) |
Not stated
|
| Runtime protection and human approval | Both built in, for any agent framework | Blocks risky prompts and behaviour; approval only for its own agents | Checks and approval rules, but only through AWS's gateway | Controls which agents may be used; approval only for ServiceNow-built agents |
| Zero-knowledge proof for your auditor | Yes |
Not stated
|
Not stated
|
Not stated
|
| Where it runs | Inside your own environment | CrowdStrike cloud | AWS cloud | Runs on the ServiceNow platform |
Also checked
Deployment
Run TraceStone on your own servers, in your private cloud or in a hybrid setup. It works with the agent framework you already use.
Start with the log collector. Move to the gateway when you are ready.
Learn more about deploymentYour environment
Put TraceStone in front of your agents' tool calls. It can block.
Call TraceStone from your agent's own code. It can block too.
Send records after the fact. Start by observing, with almost no changes.
SECURE THE FULL AI LIFECYCLE
Keep using the TraceStone products for application testing and agent security alongside runtime control and private audit.
BEFORE DEPLOYMENT · APPLICATION SECURITY
Find security gaps in AI gateways, agents and LLM applications with AI-focused testing, penetration testing and code review.
Explore vulnerability discoveryBEFORE DEPLOYMENT · AGENT SECURITY
Review an agent’s code, tools and permission model to surface prompt injection paths, privilege risks and trust-boundary issues before release.
Explore agent securityAFTER DEPLOYMENT · MONITORING
Monitor deployed agents, detect attacks and seal runtime events into tamper-evident evidence that can be checked offline.
Explore Evidence-AIDRProduct demo
Watch a reporting agent at work: safe actions pass, an export waits for your CFO, and a hidden instruction is stopped. Every action is recorded.
It checks each of your agents' actions before it runs, asks a person before risky ones, keeps signed, linked records, and gives your auditor proof that the rules held, without your customers' data.
You set the main policies. Task-aware rules come from each user request, the available tools and the workflow, and are checked against your policy and formally verified before use. High-risk actions wait for your approval.
You hand over proof, not your data. You export only the records the audit needs, so your revenue figures, prompts and replies, and request payloads stay inside your company. Zero-knowledge proofs then show that each rule held without revealing the values, for example that an action had a valid approval without showing who approved it. Your auditor checks the proofs and the signed records with an offline verifier, so they never need access to your systems.
Only the evidence the audit needs. Zero-knowledge proofs show that a rule held without revealing the value, for example that a payment stayed below its limit. Complex checks can run in a trusted execution environment (TEE). An offline verifier marks each claim as verified, failed or inconclusive; missing evidence never counts as a pass.
Each record is signed and linked to the one before it. A Merkle tree lets your auditor check one record without the whole log, and signed checkpoints go to an independent service. A removed, added or changed record is detected.
Three ways: a security gateway in front of your agents' tool calls, an embedded check in your agent's code, or a log collector that records after the fact (it cannot block, but keeps signed records). You can start with the log collector and move to the gateway later.
Nowhere new. TraceStone runs inside your environment: your servers, your private cloud or a hybrid setup. It works with any agent framework.
The MAS Guidelines on AI Risk Management (human oversight and records of AI use), the MAS SAFR paper (check each action before it runs) and the EU AI Act (logs and human oversight for high-risk AI).
With a paid pilot of 6–8 weeks on one of your agent workflows, inside your environment.
Still curious? Talk to us