New entry points
An AI assistant connected to company data is reached through an app or an API, the interface other software uses to call it. Every connection to email, files or business systems is another route an attacker can try.
AI security
Security for your AI systems
We build every AI system with controls on each request: who can use it, which data it may see, and a record of every action. Our security partner Hadrian tests what your company exposes to the internet, continuously, and confirms which weaknesses an attacker could actually use.
The risk
How AI changes security risk
AI systems add new ways into a company, and attackers now use AI themselves.
Attacks on the AI itself
Hidden instructions in a document or web page can make an AI system ignore its rules or reveal data. This is called prompt injection. The OWASP Foundation, a nonprofit that publishes widely used security guidance, ranks it first among the risks of applications built on large language models, the technology behind AI assistants.
Attackers using AI
In November 2025, Anthropic reported that a group it assessed with high confidence as Chinese state-sponsored used its AI tools to perform 80 to 90% of an espionage campaign against roughly thirty organizations. People stepped in at only a few decision points, and a small number of the attempts succeeded.
Rules that ask for evidence
The EU AI Act requires AI systems it classifies as high-risk to resist attempts to alter their use, outputs or performance. Under the EU’s NIS2 directive, management of essential and important entities must approve and oversee cybersecurity measures and can be held liable.
Sources: OWASP Top 10 for LLM Applications 2025 · Anthropic, November 13, 2025 · EU AI Act, Article 15 · NIS2 Directive (EU) 2022/2555
Built-in AI Gateway
Controls on every AI request
Security is part of the design of every AI system we build, from the first use case.
Every request to an AI model passes through the AI Gateway, a control layer that runs in your own cloud. It checks who is asking, removes personal data before anything leaves, blocks unsafe content, keeps each team within its budget and records every action in an audit trail. AI can look information up freely; any change to your systems waits for a person’s approval.
Before a system goes live, we run attack tests: thousands of attempts to trick the AI or make it leak personal data, with the results documented. The controls are mapped to the EU AI Act, the NIST AI Risk Management Framework and ISO/IEC 42001, the international standard for managing AI.
- Identity & access
- Safety guardrails
- Data privacy & redaction
- Model routing, cost & budgets
- Audit & logging
Security partner
Continuous testing
We partner with Hadrian to test your internet-facing systems the way an attacker would and to report the weaknesses it has confirmed.
Why we work with Hadrian
We work with Hadrian because its testing runs continuously, its findings are confirmed before they reach your team, and nothing has to be installed.
Hadrian was founded in 2021 by former ethical hackers: best-in-class specialists who break into systems with the owner’s permission so the weaknesses can be fixed. Its platform uses AI agents to test from the outside, with nothing to install on your systems.
Like Praexo, it works with companies in Europe and the United States.
- 01Find what is exposedHadrian continuously maps the domains, servers, certificates, cloud services and applications that can be reached from the internet, including ones nobody remembers setting up, and tests again whenever something changes.
- 02Confirm what can be exploitedAI agents check whether each finding can actually be used in an attack, so your team can focus on the weaknesses that matter.
- 03Test key applicationsFor the applications that matter most, including those through which people use your AI systems, Hadrian runs penetration tests: controlled attacks, ordered by you, on a defined app or API. Its specialists review the findings before they reach you.
- 04Fix and verifyEach confirmed weakness arrives prioritized, with the steps to fix it, in tools your team already uses, such as Jira or ServiceNow. Hadrian then tests again to confirm the fix worked. Findings are mapped to SOC 2 and ISO 27001, two common audit standards for information security, and to NIS2.
Who covers which risk
| Risk | Example | Covered by |
|---|---|---|
| Systems exposed to the internet | A forgotten test environment left online | Hadrian, continuous testing |
| Weaknesses in apps and APIs | A login flaw that shows one customer another customer’s data | Hadrian, penetration tests |
| Manipulation of the AI | Hidden instructions in a document that make the AI reveal data | AI Gateway guardrails and attack tests before go-live |
| Personal data sent to AI models | A customer file pasted into a question to the AI | AI Gateway data redaction |
| AI acting beyond its role | An AI agent changing records without approval | Access rights and human approval of changes |
| No record of AI use | Nobody can say who used AI for what | Audit trail and regular reporting |
Oversight
Reporting to leadership and the board
Executives and directors get a regular, plain-language view of AI security risk and the evidence behind it.
Regulation increasingly expects leadership to oversee cybersecurity: management approval of cybersecurity measures under NIS2 in the EU, and a description of the board’s oversight of cybersecurity risk in the annual report of US-listed companies. Each report gives leadership and the board what they need for that oversight:
- AI systems in use, what each may access and its risk level
- Internet-facing weaknesses found, fixed and still open, and how long fixes take
- Results of the latest attack tests and control reviews
- Incidents and near misses
- Decisions needed from leadership or the board
Reports are quarterly, with an immediate alert when a serious weakness is found.
Illustrative example. Not client data. AI security report
- AI systems in use
- 6 (2 high-risk)
- Internet-facing weaknesses
- 14 found · 11 fixed · 3 open
- Median time to fix
- 9 days
- Attack tests on AI systems
- 48 of 50 blocked
- Incidents
- 0 · Near misses: 1
- Decision needed
- Retire one unused customer-facing app
How we start
- 01External baselineHadrian maps what your company exposes to the internet and confirms which weaknesses matter.
- 02Review of AI systemsWe check each AI system you run or plan against the controls of the AI Gateway, our control layer for AI requests.
- 03Targeted testsPenetration tests of the apps and APIs that carry the most risk.
- 04Ongoing testing and reportingContinuous testing, a retest after each fix and regular reports to leadership and the board.
Discuss your AI security
Tell us which AI systems you run or plan. We will propose where to start and which tests are needed.