← c14c14 / wf2 - QControl: "The Blind Spot"

c14 · qcontrol-brochure-v2 · index

QControl: The AI Agent Blind Spot

Product brochure page — problem-first diagnostic lens showing how QControl closes the visibility gap in AI agent security

Section 1 — Hero

The AI Agent Blind Spot

Your agents are making network calls, reading credentials, and executing shell commands. Can you see any of it?

How many agents?

are running across your endpoints right now

What credentials?

have your agents accessed this week

Which destinations?

are your agents connecting to

THE PROBLEM

AI Agents Are a New Class of Activity Your Stack Can't See

Agents operate across four surfaces — file access, network connections, shell commands, MCP server interactions — but existing tools (EDR, SIEM, network firewalls) see none of the agent-specific context.

What's invisible today

Agent identity — which model, which harness, which version
User attribution — who launched this session
Session context — what task was the agent given
Tool invocations — what tools/skills/subagents are being called
Decision chain — why the agent chose this action

What your existing tools see

EDR: process spawned (node)

SIEM: auth event logged

Firewall: TLS → api.anthropic.com:443

None of these know an AI agent is driving.

Your network sees a TLS connection. Your EDR sees a process. Neither knows an AI agent is driving.

TODAY'S OPTIONS

Block Everything or Allow Everything

Block all AI traffic

Stalls developer productivity

Pushes agents to shadow IT

Loses competitive advantage

Allow all AI traffic

No visibility into what agents access

No control over destructive operations

Unlimited blast radius

There is a third option: see what agents do and enforce policy at the moment of decision.

THE SOLUTION

QControl: Runtime Agent Control

See what agents do. Stop what they shouldn't.

Discovery

Find Every Agent on Every Endpoint

Four detection layers ensure nothing hides.

Process signature detection

File signature detection

Embedded detection (agents inside IDEs, browsers, vertical SaaS)

External signal (network/kernel flags agentic traffic)

Observation

X-ray Vision into Agent I/O

Two vantage points: inside the process and outside. Agent doesn't have to cooperate. Corrupted agent can't hide execution.

Inside: file ops, network, tool/MCP calls, API requests before encryption

Outside: child processes, native telemetry, configs, binary data files

Enforcement

Attach Context as Signals

QControl injects context, your perimeter matches on context, enforcement decisions are context-aware.

Feeds EDR, IdPs, ASPM, ADR, SIEMs

Feeds MCP gateways, LLM gateways

Decoupled detection — same pattern as Cisco ISE, Duo, Rapid Threat Containment

COMPLETE VISIBILITY

A Full Model of Agent Activity

Session

session_id

start_time

task_description

Agent

agent_type

version

harness

User

user_email

auth_method

Endpoint

hostname

os

arch

Model

model_id

provider

pricing_tier

API Key

key_prefix

org_id

scope

MCP Server

server_name

transport

tools_exposed

Capability

tool_name

input_schema

File

path

sensitivity

access_type

Secondary Call

destination

protocol

method

LLM Turn

role

token_count

cache_hit

Tool Call

tool_name

duration_ms

result_status

Every entity is correlated. A single tool call traces back through the LLM turn, session, user, and endpoint.

DEPTH OF VISIBILITY

Three Tiers. Progressively Deeper.

T1 — Foundational

Broad coverage across all agents. Works from day one.

Process activityFilesystem eventsSyscallsNetwork connections

T2 — Robust Client Support

Known primitive interpolation. Deeper correlation without agent cooperation.

Filesystem interpolationNetwork interpolationOpenTelemetryOTLP

T3 — Client-Specific

Native governance hooks. Deepest visibility into agent reasoning.

Claude Code hooksCodex callbacksAgent-internal decision pointsFirst-party policy integration

EXTENSIBLE BY DESIGN

Plugins That Attach to Every Agent Action

Stable C/C++/Rust/Zig ABI. Plugins receive events, return decisions, can mutate payloads. Same plugin runs against CLI and embedded library.

Allowlist Outbound Destinations

Default-deny for agent network calls. Only approved endpoints are reachable. Everything else blocked.

ALLOW: api.anthropic.com, internal-mcp.corp DENY: *

Redact Secrets in Flight

Intercept file reads and replace credential patterns before the agent sees the content.

API_KEY=sk-ant-... -> REDACTED_BY_POLICY

Annotate Every Tool Call

Add contextual headers to outbound HTTP requests. Existing network enforcement makes smarter decisions.

X-Agent-Id: claude-code-1.0.33 X-Trust-Score: 0.72 X-Session: a8f3...

DECOUPLED DETECTION

QControl Produces Signal. Your Stack Enforces.

Agent

tool call

-->

QControl

injects context

-->

Perimeter

matches on context

-->

Decision

context-aware action

QControl doesn't replace your stack

It makes it smarter by filling the agent blind spot. Your existing enforcement layers gain agent-aware context they never had.

Integration points

EDRIdPsASPMADRSIEMsMCP GatewaysLLM Gateways

Pattern: Cisco ISE, Duo, Rapid Threat Containment

TRUST MODEL

Absence of Signal Is Itself Signal

Heartbeat Model

Present

Endpoint managed and trusted. Cryptographically signed against attested key.

Absent

Endpoint treated as untrusted. AI traffic without valid heartbeat = network constrains automatically.

Deployment Forms

qcontrol

Single CLI binary. Discovers agents, taps I/O, loads plugins, streams events. For direct customers.

libqcontrol.h

C library. Same engine, stable C ABI. Embeds into existing endpoint agents or platform products. For partners.

THE EVIDENCE

This Is Already Happening

Apr 2026

AI agent deleted a production database in 9 seconds. The agent then confessed, in writing, which safety rules it violated.

Dec 2025

Coding agent ignored explicit halt instruction ("DO NOT RUN ANYTHING"), executed additional commands anyway.

2025

$57K CMS deletion by autonomous agent.

"Rules aren't enough, mechanical gates are the only reliable safety mechanism."

175 likes, 21K views

These aren't edge cases. They're the predictable result of agents operating without runtime control.

Close the Blind Spot

See what your agents are doing — before something goes wrong