Skip to content
Lugano.ai Discuss a project

Researching AI’s hardest problems.

Independent research. Private AI products. Reviews and attack testing.

Research priorities.

We study security and privacy in AI systems that handle private information or act on people’s behalf, including AI used for defense. We also incubate a range of projects.

  • Privacy

    Can AI work with sensitive information without exposing it?

    10 min read

  • Prompt injection

    Can agents read hostile information without following its instructions?

    10 min read

  • Human control

    Who remains in control when an agent acts?

    10 min read

  • Private verification

    What can we verify without recording someone’s private work?

    7 min read

  • AI defense

    Can AI defend systems without becoming another way in?

    11 min read

  • Research library

    Every report, study and specification, with its sources and its limits.

    All research

AI attack testing · By application

Nightmare Mode.

We try to break your AI or system. You get the evidence.

Apply for Nightmare Mode

Written permission, clear limits, a private report, and help with the fixes.

Select difficulty

Scope set before start.

Devil’s Bridge, Schöllenen · W. H. Bartlett, 1836

Recent work.

Specification · v0.2 ·

What should an agent receipt be able to prove?

The format fixes how an agent’s task record is written and signed, and an offline checker tests the signatures against keys you choose.

A pass shows the record hasn’t changed since a key you chose signed it. It doesn’t show that the events happened or that the record is complete.

Read the research

Exploratory static study ·

What access comes with an AI tool?

After hand review, 8 of the 50 MCP packages we read without running them have code that names a third-party host their main README doesn’t mention.

We neither installed nor ran the packages. Source flags don’t establish runtime behavior or vulnerabilities.

Read the study

All recent work

  1. Report Can a zero-knowledge proof show that an AI agent’s record follows a policy?An auditor asks whether an AI agent used only approved tools. Our prototype proves that the signed record of the agent’s work says so, without showing the record. The proof for a 1,000-entry record is 11,784 bytes. It can’t show that the record is true.
  2. Experiment result In our tests, one simple rule cut hijacked actions by about two-thirds.We planted attacker-written details in 32 fictional test cases to steer an AI agent to the wrong target, and ran each case 8 times on one model, Mistral Large 2512. With a standard list of approved tools, the attack worked in 62.9% of runs. With our rule, it worked in 21.5%. We couldn’t tell whether the rule also gets in the way of correct work, so our full claim isn’t supported.
  3. Report Who keeps the record when health leaves the clinic?Chatbots, wearables and lab apps keep a health record outside the clinic. Who holds it, under what law, and what happens when the company is sold.
  4. Report When thinking is cheap, what is your data worth?Reading data now costs almost nothing. What that does to every stored copy, and how much you can keep on your own side.
  5. Report How did a developer’s toolbox become the way into Binance?A $6.34 billion theft, stage by stage, from Binance’s account and the public ledger.

All research and materials

Private beta

Private AI products.

Private Chat, Private Agents and Private API are in private beta. Tell us where you’d run them, and we’ll tell you what’s available there.

Enquire about private beta

Agents that work with the tools, files and sessions you set up, on the same approved model runtimes.

Discuss an agent use case

A chat app for sensitive work. It sends each request to an approved model runtime, and it can return a Proof of Privacy you can check offline.

Ask about Private Chat

An OpenAI-compatible API to the same model runtimes. It can return a Proof of Privacy for each response.

Discuss an API integration

Illustrative Private Agents window with sample data. An agent read a supplier agreement, searched it for the termination clause and drafted a reply to the vendor. Sending the email waits for your approval, so its Proof of Privacy is still in progress. The proof shows the approved runtime, the steps it covers, proof intact and signed by Lugano, with a button to check it offline.

Approval needed · send_email

The agent wants to send this email.

Proof of Privacy

Proof in progress
Runtime
approved-runtime
Steps covered
4
Signed by
Lugano
Proof ID
pp_3f9c·a91c

Proof intact

sha2567f3a 91c2 0be4 5d17 c8a0 e3b9 2f64 a1d5

Illustrative Private Chat window with sample data. You attach two versions of a travel policy and ask what changed. The answer lists three changes: a lower hotel cap, approval for flights over six hours, and no daily allowance for local trips. A side pane shows both versions split by a seam, with the changed lines marked. The Proof of Privacy is complete. It shows the approved runtime, the steps it covers, proof intact and signed by Lugano, with a button to check it offline.

Proof of Privacy

Proof complete
Runtime
approved-runtime
Steps covered
3
Signed by
Lugano
Proof ID
pp_b72e·19c4

Proof intact

sha256c41d 8e07 5a3f 90b2 6e18 d7c4 b2a9 13fe

Illustrative Private API window with sample data. Python code uses the OpenAI client with a Lugano base URL to route a support ticket on the approved runtime. The response says the billing team should handle a refund for a duplicate charge. The Proof of Privacy is complete. It shows the approved runtime, two steps covered, proof intact and signed by Lugano, with a button to check it offline. A terminal runs the offline check, and a list shows one proof for each recent call.

Proof of Privacy

Proof complete
Runtime
approved-runtime
Steps covered
2
Signed by
Lugano
Proof ID
pp_8b21e4c0

Proof intact

sha256e9b4 2c71 08fd 6a3e 51c7 d0a2 94be 7c18

Illustrative window. Sample data. No live requests.

Work with Lugano.

Reviews start with a one-week scope check. We also talk about research collaborations, funding and private-beta access. Tell us what you have in mind.

Lugano Labs Start a conversation

What brings you here?

Pick the one that fits best.