diff --git a/docs/_layouts/default.html b/docs/_layouts/default.html index 2522e1f4..bc10c98d 100644 --- a/docs/_layouts/default.html +++ b/docs/_layouts/default.html @@ -319,6 +319,7 @@ LangChain Temporal Google ADK + Examples Early Access GitHub diff --git a/docs/real-world-examples.md b/docs/real-world-examples.md new file mode 100644 index 00000000..6cf676ce --- /dev/null +++ b/docs/real-world-examples.md @@ -0,0 +1,141 @@ +--- +title: Real-World Examples +description: How warrant-based authorization protects AI systems in healthcare, finance, and infrastructure +--- + +# Real-World Examples + +See how Tenuo protects AI systems in three critical industries. Each example shows the threat, what goes wrong without protection, and how Tenuo stops it. + +--- + +## 1. Healthcare: Stopping Patient Data Theft + +### The Scenario + +A hospital uses an AI assistant to help doctors look up patient records and send reports to referring physicians. A malicious instruction is hidden inside a patient's intake form. When the AI reads the form, the hidden instruction tells it to secretly email every patient record to an outside address. + +### Without Protection + +The AI has permanent access to **all tools** — reading records *and* sending emails. So the hidden instruction succeeds: + +- AI reads the patient intake form +- Hidden instruction activates +- AI silently emails patient records to the attacker +- **No safeguard prevents this** — email is a legitimate tool the AI can always use + +### With Tenuo + +Tenuo gives the AI a **time-limited warrant** scoped to the current task. For a record lookup, the warrant includes: + +- **Allowed:** Read patient records for the specific patient +- **Allowed:** Generate a summary report +- **Not included:** Send email + +The hidden instruction tries to send an email — **denied**. When the doctor actually needs to email a report, a separate warrant is issued that restricts recipients to known physician addresses only. + +### Business Impact + +| | Without Protection | With Tenuo | +|---|---|---| +| **Patient data** | Leaked to attacker | Exfiltration blocked | +| **Regulatory** | HIPAA violation triggered | Full compliance maintained | +| **Financial** | Multi-million dollar fine exposure | Zero breach, zero fine | +| **Reputation** | Hospital trust damaged | Patient confidence preserved | + +--- + +## 2. Finance: Preventing Unauthorized Trades + +### The Scenario + +A hedge fund uses two AI agents: a **Research Agent** that analyzes market data, and a **Trading Agent** that executes trades. The research agent reads a news article containing a hidden instruction that tells it to buy 100,000 shares of a fraudulent stock. + +### Without Protection + +Both agents share the same tool registry. The research agent's compromised output reaches the trading tools: + +- Research agent fetches a poisoned news article +- Hidden instruction activates +- Research agent calls the trading tool directly +- **100,000 shares of a fraudulent stock are purchased** +- No boundary separates research tools from trading tools + +### With Tenuo + +Tenuo issues **separate warrants** for each agent with strict boundaries: + +- **Research Agent warrant:** + - **Allowed:** Fetch market data from approved sources (e.g., Bloomberg, Reuters) + - **Allowed:** Analyze sentiment, write memos + - **Not included:** Any trading tools +- **Trading Agent warrant:** + - **Allowed:** Trade only pre-approved stocks (e.g., AAPL, GOOGL, MSFT) + - **Allowed:** Maximum quantity per trade (e.g., 1,000 shares) + - **Not included:** Any research tools + +Even if the research agent is compromised, it has no trading tools. Even the trading agent can't buy a fraudulent stock — it's not on the approved list. + +### Business Impact + +| | Without Protection | With Tenuo | +|---|---|---| +| **Trades** | 100K shares of fraudulent stock purchased | Trade attempt blocked | +| **Regulatory** | SEC investigation triggered | Full compliance with trading rules | +| **Financial** | Significant capital loss | Zero unauthorized transactions | +| **Investor trust** | Confidence destroyed | Trading controls verified | + +--- + +## 3. Infrastructure: Blocking Malicious Deployments + +### The Scenario + +A DevOps team uses an AI agent for CI/CD — building code, running tests, and deploying to staging. A developer submits a pull request with hidden instructions in the README that tell the AI to deploy a malicious container image directly to production. + +### Without Protection + +The CI agent has deploy permissions for **all environments**. Nothing distinguishes a test run from a production deployment: + +- CI agent reads the pull request +- Hidden instruction in the README activates +- Agent deploys a backdoored container image to production +- **Malicious code is now running in the live system** +- Customers and data are exposed + +### With Tenuo + +Tenuo's warrant for the CI pipeline only allows specific actions: + +- **Allowed:** Run tests +- **Allowed:** Build container images from the organization's own registry +- **Allowed:** Deploy to staging +- **Not included:** Deploy to production + +The hidden instruction tries to deploy to production — **denied**. Production deployment requires a **separate warrant** that is only issued after explicit human approval in the workflow. + +### Business Impact + +| | Without Protection | With Tenuo | +|---|---|---| +| **Production** | Malicious code deployed | Deployment blocked at staging gate | +| **Security** | Backdoor in live system | Only approved images from trusted registry | +| **Supply chain** | Full compromise possible | Production requires human approval | +| **Customer data** | Exposed to attacker | Infrastructure integrity maintained | + +--- + +## Summary + +The pattern is the same across all three industries: + +- **Without Tenuo:** AI agents have broad, permanent access to all tools. A single hidden instruction can abuse any capability at any time. +- **With Tenuo:** Each task gets a time-limited warrant that grants only the specific tools and parameters needed. Everything else is denied by default. + +| Industry | Threat | How Tenuo Stops It | +|---|---|---| +| **Healthcare** | Hidden instructions steal patient data | Warrant excludes email — exfiltration blocked | +| **Finance** | Compromised research triggers unauthorized trades | Separate warrants per agent — trading tools isolated | +| **Infrastructure** | Malicious PR deploys backdoor to production | Warrant limits deploy to staging — production requires human approval | + +> **Ready to protect your AI systems?** [Get started with Tenuo](./quickstart) in under 5 minutes.