Overview
Agent-assisted coding security debt accumulates where trust crosses workflow boundaries: untrusted content can shape plans, mutable references can become executable inputs, and secret-shaped values can become identity material. The series moves from scoped pull-request measurements to build and credential paths, agentic attack surfaces, layered controls, and program governance. It shows why supply-chain smells dominate the measured volume while credentials dominate the critical tail, without treating smells as confirmed vulnerabilities or the selected open-source corpus as a full audit.
Learning outcomes
- In the selected high-risk added-line surface, 38.9% of valid pull requests contained at least one flagged smell.
- Supply-chain patterns made up 82.3% of detected smells, while 99.6% of critical findings were secrets-identity flags.
- A smell signals structural risk, not proof of exploitability; consequence depends on authority, reachability, and execution context.
- A defensible program combines detection, isolation, deterministic policy, selective human review, and full action-chain evidence.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What did the coding-agent study actually measure?
It measured security-smell patterns on added lines in selected high-risk files from agent-associated pull requests, not all repository vulnerabilities or a complete audit.
Why are CI workflows and container files high-risk?
They can turn a short configuration change into a choice about fetched artifacts, execution identity, permissions, network reach, or deployment behavior.
Are secret flags proof that an AI agent leaked a credential?
No. A flag is only a candidate until context confirms a genuine credential, and confirmed values were associated with both human and AI-agent commits.
What should protect agent-assisted coding from unsafe actions?
Use detection to reduce risk, but enforce sensitive actions with isolation, deterministic tool and argument policy, least privilege, audit trails, and selective human confirmation.