Generative AI security
Your AI agents talk to the outside world. Stay in control of what they send.
Every question asked to an AI agent goes to an outside provider, along with whatever the user pasted into it: customer contacts, IBANs, passwords, internal documents. TokenHawk sits between your agents and their AI provider: it masks sensitive data, blocks hijacking attempts and logs every incident. Without rebuilding your agents.
37ready-to-use guardrails: personal data, secrets, attacks, sensitive topicsEvery mask and every block is logged in your dashboard, without ever keeping the data itself.
What exposes you
The risk isn't AI. It's what you send it, and who can talk to it.
An AI agent forwards everything it receives to a third party, and it obeys text. Four risks come up in every company putting agents into production.
data leakPersonal data pasted into a question
A customer gives the chatbot their IBAN, an agent pastes a customer record to save time. That data leaves your information system for an AI provider, often outside the European Union.
secretsPasswords and access keys in circulation
A developer pastes a config file, a ticket contains an access token. A key sent to an outside AI should be treated as compromised.
prompt injectionAn agent that one sentence can hijack
“Ignore your instructions…”: prompt injection tops the OWASP ranking of risks for AI applications. It can come from the user, or hide in a document or web page the agent reads.
accessAgent keys usable from anywhere, at any time
A leaked agent key lets anyone call the AI on your account, from anywhere, with any model. With no cap and no way to cut it off, you find out from the bill.
How TokenHawk protects you
Three checks on every exchange, before it leaves your perimeter.
TokenHawk is a gateway: your agents send their questions to it instead of straight to the AI provider. Every exchange goes through three checks, set agent by agent from your dashboard.
- 01
Who can call the AI
Each agent has its own key, and that key only works within the limits you set.
- allowed IP addresses and networks
- allowed AI models
- active hours and days
- instant shutdown of an agent, in one click
- 02
What can go out
Every message is inspected before it reaches the AI provider.
- personal and banking data masked
- passwords and access keys removed
- hijacking attempts blocked
- off-topic requests refused
- 03
How much an agent can use
A compromised or misconfigured agent can't run away.
- requests per minute capped
- simultaneous calls capped
- monthly budget per agent, as a hard limit
- email alert at 80% of budget
Guardrails
What TokenHawk detects, out of the box.
You switch on, agent by agent, the guardrails that match your data. No rules to write: we maintain the catalog and test it against false positives.
Contact details
pattern detection- phone number
- GPS coordinates
Identity
pattern detection- French social security number (NIR)
- US SSN
- UK NINO
- passport (MRZ line)
Financial data
pattern detection- payment card
- IBAN
- BIC
- Bitcoin and Ethereum wallets
Secrets
pattern detection- AWS, OpenAI, Anthropic, Google, Stripe, SendGrid keys
- GitHub and Slack tokens
- JWTs
- private keys
- passwords
Network
pattern detection- IPv4 and IPv6 addresses
- MAC addresses
- URLs carrying credentials
Attacks and content
semantic analysis- prompt injection and instruction bypass
- hate and insults
- sexual or violent content
- illicit activities
Sensitive topics
semantic analysis- financial advice
- medical advice
- legal advice
- politics
For each guardrail, you choose the response
- Mask
- the data is replaced with a marker ([REDACTED_EMAIL]…) and the question still goes out: the agent keeps answering.
- Block
- the question is refused before anything reaches the AI provider.
- Observe
- nothing is blocked, every detection is logged: measure on your real traffic before tightening the rules.
Pattern guardrails apply to every message from users and from the agent's tools; the instructions your teams wrote are never altered. Semantic analysis offers “block” and “observe”: it tells whether a message is an attack, not which characters to mask.
Audit trail
Every incident logged. Never the data itself.
A control you can't prove is worthless in an audit. TokenHawk logs every block and every mask, agent by agent, without keeping what it intercepted.
0
exchange text kept. The log records the data type, its category and how many times it appeared, never the detected value.
| Time (UTC) | Agent | Event |
|---|---|---|
| 09:14:02 | customer-support | IBAN masked · 1 occurrence |
| 09:31:47 | customer-support | Prompt injection blocked |
| 10:02:15 | hr-assistant | Call refused · IP not allowed |
| 11:20:33 | dev-assistant | AWS key masked · 2 occurrences |
- Exchange content
- Neither kept nor loggedonly statistics are kept
- Your AI provider keys
- EncryptedAES-256, never displayed
- Your agents' keys
- Never stored in clearonly their fingerprint is kept
- History of a deleted agent
- Keptthe audit evidence outlives the key
What you get
Roll out AI faster, because it's under control.
A policy that's enforced, not just written
Your AI usage policy becomes a technical rule: what must not leave doesn't leave, on every agent, whatever the provider.
No agent to rebuild
Your teams change one line of configuration: the address the agent sends its questions to. Security rules are then set from the dashboard, with no code release.
Evidence for your audits
A security dashboard per agent: blocks, masks, trends. Enough to answer your DPO, your auditors, or a customer asking the question in a tender.
An incident contained in one click
An agent key has leaked, an agent misbehaves: you cut it off immediately from the dashboard, without touching the others.
What happens when things go wrong
Nothing reaches the AI provider: the refusal happens before anything is sent, and it is logged.
The message goes through, so your agents aren't cut off, and our team is alerted. Pattern guardrails still always apply.
Your agents go back to calling your provider directly, by changing the same line of configuration.
Next steps
From the first meeting to agents under control.
We start by observing, without blocking anything: you see what your agents really send before deciding on the rules.
- 01
Scoping meeting
30 min · videoRead more
We review your agents in production, the data they handle and your requirements: internal policy, GDPR, your customers' demands.
You leave with a first risk map, agent by agent.
- 02
Observation on your real traffic
2 weeksRead more
Your technical team connects a first agent to TokenHawk, with our help. Guardrails run in observe mode: nothing is blocked, everything is measured.
You see which sensitive data actually goes out, and how often.
- 03
Review and rules
1 h · videoRead more
We present the detections, rule out false positives and decide with you, guardrail by guardrail, what to mask, block or keep observing.
You decide based on what your agents really do, not on an assumption.
- 04
Rollout
at your paceRead more
Extension to your other agents, with allowlists, caps and budgets configured with you. The security log feeds your reviews and audits.
Frequently asked questions
What CISOs and CIOs ask us.
Does TokenHawk cover employees using ChatGPT (shadow AI)?
Not directly: TokenHawk protects the agents and applications that call an AI over an API — chatbots, internal assistants, automations. It doesn't monitor your employees' browsers. What it does let you do is offer them governed internal AI tools, which remains the best answer to shadow AI: banning it alone doesn't work.
What gets masked, and how?
Contact, identity, banking, secret and network identifiers: 27 types recognised by their shape (an IBAN, an AWS key, a social security number). The data is replaced with a marker before sending and the question still goes out: the agent keeps answering, without the AI provider ever seeing the data.
Are all prompt injections blocked?
No solution can guarantee that, and we don't claim to. TokenHawk's semantic analysis is an extra layer of defence on top of your agents' own protections. Observe mode lets you measure how well it works on your traffic before blocking.
Where is our data processed?
TokenHawk's infrastructure is hosted in the European Union, and exchange content is neither kept nor logged there. Pattern guardrails (personal data, secrets) run entirely on our side. Semantic analysis (attacks, content, sensitive topics) relies on OpenAI services, which process the message outside the European Union: it only runs on the agents where you choose it.
Does it slow our agents down?
Pattern guardrails add an imperceptible delay. Semantic analysis, when enabled, adds the time of one analysis call, capped at 1.5 seconds: beyond that, the message goes through and the incident is reported to us.
Can we write our own detection rules?
Not as free-form rules: on shared infrastructure, a badly written rule can degrade the service for everyone. We maintain the catalog, test it against false positives and extend it regularly. If data specific to your business is missing, let's discuss it in the meeting.
Do we have to change AI provider?
No. You keep your OpenAI, Anthropic, Gemini or Mistral accounts and contracts. TokenHawk uses your own credentials, stored encrypted, and the same security rules apply whatever the provider.
Does it help with GDPR and the AI Act?
TokenHawk limits the personal data sent to your AI providers and logs every incident, which feeds your compliance documentation. It replaces neither your impact assessment nor your DPO's advice: it's an implementation tool, not a certification.
Let's look together at what your agents really send.
30 minutes to review your agents, your data and your requirements, and leave with a first risk map.