Signed Tools

Mathematical certainty for agentic execution

Traditional AI security relies on probabilistic firewalls and brittle hooks to guess whether an agent is doing something malicious. But agents have access to your databases, APIs, and cloud infrastructure, which makes prompt injection a remote code execution vulnerability.

A cryptographically signed tool executes while an unsigned tool is rejected at runtime

NuDay takes a different approach: mathematical certainty. We secure autonomous workflows by cryptographically signing the agent's tools, guardrails, and instructions.

If a command doesn't have the math to back it up, it simply doesn't run.

The risk

When prompt injection becomes command injection

When an agent reads a poisoned web page, untrusted PDF, or malicious email, an indirect prompt injection can overwrite its system prompt, hijacking the LLM and instructing it to run a backend tool, such as exporting a database or emailing a secure file to an external server.

Legacy systems try to stop this with brittle defenses:

Prompt firewalls

Scan text for malicious intent before the LLM reads it. Easily bypassed by encoding or novel attack phrasing.

Execution hooks

Use Python if/else statements to intercept the action. You are relying on software to police an unpredictable neural network.

If the attacker tricks the LLM or bypasses the filter, the rogue tool executes.

What we sign

What NuDay cryptographically signs

We physically separate the AI's reasoning (the LLM) from its execution environment using Public Key Infrastructure. Here is exactly what we sign and lock down:

1

Agent tools

What it is

The Python scripts, API calls, and functions your agent uses to interact with the world.

How it's protected

Every approved tool is hashed and signed with an enterprise private key. When the LLM requests a tool, the NuDay runtime verifies the cryptographic signature before execution.

The protection

Zero-click RCE and supply-chain attacks. If a hijacked prompt injects a rogue tool, or a compromised library alters the payload, the signature fails and execution is blocked instantly at runtime.

2

Behavioral guardrails

What it is

The hard boundary conditions that dictate what an agent cannot do, such as never expose PII, or never execute trades over $10k without human approval.

How it's protected

Guardrails are signed and immutably bound to the agent's core identity.

The protection

Instruction-override attacks. Because guardrails are signed at the infrastructure layer, not just typed into the system prompt, the LLM is physically incapable of overriding or forgetting its security boundaries.

3

Operational guidelines

What it is

The step-by-step standard operating procedures an agent must follow to complete a task compliantly.

How it's protected

Guidelines are version-controlled and signed.

The protection

Behavioral drift and hallucination. If the LLM hallucinates an unapproved step in a critical workflow, the runtime rejects it: the hallucinated path lacks the approved guideline's signature.

4

Agent skills and identity

What it is

The verified capabilities assigned to a specific agent, such as a read-only skill for the HR database.

How it's protected

Skills are cryptographically bound to the agent's OIDC identity token.

The protection

Lateral movement and privilege escalation. A compromised low-level agent cannot impersonate a high-level agent or grant itself new skills; the infrastructure demands cryptographic proof of the assignment.

Comparison

The state of agentic security

Many vendors claim to secure agents; almost all rely on probabilistic filtering or fragile code-level hooks. NuDay is the only platform enforcing deterministic, cryptographic execution.

Security capabilityNuDayAI firewallsAuth layersOrchestrators
Core security mechanismCryptographic signatures (PKI)Input/output prompt filteringOAuth and identity mappingPython if/else hooks
Protects against zero-click RCEYes (math-based blocking)Partial (relies on detecting intent)No (focuses on user auth)No (developer-written filters)
Digitally signed guardrailsYes (tamper-proof)No (text-based wrappers)N/ANo (system prompt only)
Execution determinism100% deterministic verificationProbabilistic (LLM detection)Deterministic (OAuth scopes only)Probabilistic (bypassable)
Supply-chain tool protectionYes (verifies tool hash)NoNoNo
Best used forRegulated enterprise deploymentsGeneral chatbot safetyStandardizing SaaS integrationsRapid prototyping

Secure the execution layer

Mathematical certainty for autonomous AI.

See how NuDay's cryptographic signatures give deterministic, tamper-proof control over what your agents can execute.