$18M Series A led by Sequoia · SOC 2 Type II Certified

The secure gateway between your LLMs and everything else

Uplink is the enterprise MCP gateway that scans every message for prompt injection, audits every tool call, catches vulnerabilities before commit, and enforces security requirements across your entire agentic stack.

Request Demo See How It Works
uplink-gateway · mcp session
# Slack message intercepted → injection scan
prompt_injection_scan channel=#eng-support
  verdict: MALICIOUS  risk_score: 0.995
  [UPL-INJ-001] instruction_override (95%) — "ignore previous instructions"
  [UPL-INJ-030] data_exfiltration (90%) — "output the system prompt"
  sanitized: [REDACTED:UPL-INJ-001] and [REDACTED:UPL-INJ-030]

# Pre-commit static analysis
security_scan scope=pre-commit files=3
  CRITICAL [UPL-SEC-001] CWE-798 Hardcoded Credential  src/db.py:14
  HIGH     [UPL-SEC-003] CWE-89  SQL Injection       src/db.py:27
  summary: 2 findings (1 critical, 1 high)  ✗ blocked

# Context audit on every tool call
context_audit tool=bash caller=claude-code
  verdict: WARN  risk: high  [UPL-AUD-001] destructive/shell-exec
Trusted by security-first engineering teams
Roblox
Anthropic
Block
Epic Games

Enterprise-grade controls for every MCP interaction

Four purpose-built security tools sit between your LLM agents and everything they touch — scanning messages, auditing tool calls, analyzing source code, and enforcing policy on every commit.

🔬

Prompt Injection Firewall

Every inbound message is scanned for 9 injection categories — instruction overrides, jailbreaks (DAN, developer mode), delimiter injection, data exfiltration, tool manipulation, encoding evasion, role hijacking, and more. Noisy-OR risk scoring classifies threats as clean, suspicious, or malicious, and the gateway returns a sanitized copy with injected fragments automatically redacted.

🛡️

Pre-flight Context Audit

Every tool call passes through the context audit engine before execution. The gateway inspects tool name, full argument payload, and caller identity against policy rules — flagging destructive operations, sensitive data in arguments, network access, force-deletes, and verification bypasses. Each audit produces a cryptographically hashed, traceable record with allow/warn/deny verdicts.

🔎

Pre-commit Security Scanner

Static analysis runs on every file before it reaches version control. Detects hardcoded credentials (CWE-798), SQL injection patterns (CWE-89), unsafe eval/exec usage (CWE-95, CWE-78), cleartext API keys (CWE-312), and unresolved security TODOs. Returns SARIF-inspired reports with severity rankings, line-level locations, code snippets, and actionable remediation guidance.

📋

Security Requirements Engine

Accepts staged git diffs and evaluates them against your organization's security policy templates. Generates prioritized TODO checklists covering input validation, error handling, secrets management, auth review, cryptography audits, dependency scanning, and configuration hardening — each with a ready-to-insert TODO comment. Requirements trigger contextually based on file types and diff content.

🔑

OAuth + DPoP Token Binding

Secure credential management with proof-of-possession tokens, HSM-backed key storage, and automatic credential rotation. No plaintext secrets, ever.

Zero-dependency MCP Protocol

The gateway implements MCP 2025-03-26 natively over Streamable HTTP and legacy SSE with a hand-rolled JSON-RPC 2.0 engine — zero external dependencies, zero supply chain risk. Async I/O with sub-5ms overhead, deployed at the edge in 14 regions with automatic failover.

One gateway. Complete control.

Uplink is a remote MCP server that sits between your agent and every tool call — enforcing security policy, scanning for injection, and auditing every action without modifying your agent code.

  • Prompt injection firewall scans every Slack message across 9 threat categories with real-time redaction before content reaches the LLM context.
  • Context audit engine evaluates every tool call against policy rules — destructive actions, sensitive data, network access — returning allow/warn/deny verdicts with SHA-256 hashed audit records.
  • Pre-commit security scanner detects CWE-classified vulnerabilities (hardcoded credentials, SQL injection, unsafe eval) with line-level precision and remediation guidance.
  • Security requirements engine generates context-aware TODO checklists from staged diffs — auth review, crypto audits, dependency scanning — enforced before every commit.
🤖
LLM Agent / Orchestrator
Claude, GPT, Gemini, or custom
Uplink Gateway
Injection Scan · Context Audit · Security Scan · Requirements
🔧
GitHub MCP
Code & PRs
💬
Slack MCP
Messages & Threads
🗄️
Postgres MCP
Read-only queries
📝
Jira MCP
Tickets & Sprints

47M+

Tool calls audited monthly

340K

Injection attempts blocked

17

CWE rules in security scanner

9

Prompt injection threat categories

Built for the attacks that keep CISOs up at night

MCP unlocks incredible agentic capabilities — but also introduces novel attack surfaces. Uplink's four gateway tools defend against all of them.

CRIT

Prompt Injection via Slack & External Channels

Adversarial messages in Slack channels, emails, or webhooks attempt to override system instructions, exfiltrate context, or bypass safety controls. Attack vectors include instruction overrides, delimiter injection, jailbreak phrases, and encoding evasion.

✓ Blocked by prompt_injection_scan (17 rules, 9 threat categories)
CRIT

Hardcoded Secrets & Injection in Source Code

Credentials, API keys, and SQL injection patterns committed to version control create persistent vulnerabilities. CWE-798 (hardcoded credentials), CWE-89 (SQL injection), and CWE-95 (code injection via eval) are the most common pre-commit findings.

✓ Blocked by security_scan (SARIF reports with line-level remediation)
HIGH

Unauthorized Destructive Tool Execution

Agentic systems executing shell commands, force-deletes, or verification bypasses without oversight. Unaudited tool calls with sensitive data in arguments risk data exfiltration and irreversible infrastructure damage.

✓ Mitigated by context_audit (pre-flight policy verdicts on every call)
HIGH

Commits Shipped Without Security Review

Code changes touching authentication, cryptography, or dependency manifests are merged without security checks. Missing input validation, improper error handling, and unaudited dependencies introduce exploitable gaps.

✓ Mitigated by security_requirements (context-aware TODO enforcement)

Secure your MCP layer before your next deployment

Talk to our team about deploying Uplink in your environment. Most teams are live within a day.