How we hardened an agentic software stack, from token sprawl to agent-ready APIs

Agentic AI moves fast, until it needs to touch real systems. Fleets of agents write code and call APIs, multiplying the number of things that ship and the credentials that authorize them. This team was shipping AI features at speed, but the security and governance underneath hadn't kept up, and a live supply-chain attack made that impossible to ignore.

The Challenge

AI velocity had outrun the platform's controls. Three problems compounded each other, then an incident forced the issue:

1. 01

   Token & secret sprawl. Personal access tokens, API keys, service accounts and agent credentials outnumbered people, long-lived, over-scoped and rarely rotated.

2. 02

   APIs not agent-ready. Agents could reach systems they shouldn't and stall on ones they should, there was no clean, governed layer for them to act through.

3. 03

   Unevaluated, unhardened delivery. Human- and agent-written code shipped without evals or supply-chain controls, the exact gap a live GitHub supply-chain attack then exploited, abusing a developer token to push malware.

The agents were productive. The problem was everything they could touch, and the fact that a single abused token could turn that reach against us. The engagement brief

The Solution

Periscope tracked, contained and remediated the active attack, then closed the gaps behind it: governing non-human identities, building an agent-ready API and pipeline layer, and wrapping every AI feature in evaluation and cost control.

Incident response first. We tracked the abused developer token, contained the malware built to scan code and artifacts, mine crypto and encrypt files, and remediated the blast radius.

Non-human identity governance. A read-only exposure scan across repos, CI, cloud and SaaS, age, scope, last use, leakage, then rotation, scope reduction, short-lived credentials and workload identity federation.

Agent-ready APIs & MCP. A clean, governed layer, APIs and MCP servers, so agents act only on what they should, with the right controls on every endpoint.

Pipelines for humans and agents. Build pipelines both people and agents move through across environments, with cloud workloads sized for agent and API traffic.

Evals & cost control. Eval suites, drift monitoring and token telemetry with smart model routing, so AI features stay accurate and their economics are known.

The Results

The attack was contained and the gaps behind it closed, turning a fast-but-fragile AI build into one that could scale safely.

Attack contained & remediated. The active supply-chain compromise was tracked, stopped and cleaned up, with every credential in reach rotated and scoped down.

Governed non-human identities. Tokens, keys and agent credentials moved to short-lived, least-privilege access under continuous monitoring for abuse.

Agents that act safely. An agent-ready API and pipeline layer let the team keep shipping agentic features, without agents reaching what they shouldn't.

Speed that compounds. With evals and cost control in place, AI velocity stopped accumulating security and cost debt behind it.

We didn't slow the team down, we made the platform underneath them safe to move fast on: governed identities, agent-ready APIs, and a pipeline that catches the next attack. Outcome summary