Episode 35: AI is transforming Security

Published: Tuesday, Jun 9, 2026 • Duration: 61 minutes • Season 1

Supply chain attacks, AWS CodeArtifact, agentic workflows, auto mode sandboxing, and local memory systems with Hermes agent.

Watch on YouTube

`summarize "https://youtu.be/W7QcWP8VngI" --timestamps --slides`

An informal conversation about how AI is reshaping security and developer operations. Topics range from supply chain attacks and private registries to agentic workflows, auto mode sandboxing, and local memory systems.

Supply chain attacks & namesquatting

Red Hat GitHub breach, malicious packages propagating through frameworks and VS Code plugins, and PNPM becoming aggressive about disabling install-time hooks.

AWS CodeArtifact replacing Nexus

IAM/SSO integration makes rollout simple: read-only scopes for developers, GitHub Actions OIDC for publishing, pull-through caching against public registry outages — versus the cost and complexity of JFrog/Artifactory enterprise features.

AI reshaping the security profession

Security checklists become automated tests and guardrails; agents execute compliance checks across AWS configs and Git repos. The profession shifts from spreadsheet auditors to engineers who define what “good” looks like.

Agentic workflows: dynamic workflows & auto mode

Auto mode delegates command approvals to a model-based classifier — 0.4% of benign commands blocked, 17% of risky actions got through. Docker sandboxes and Cowork isolation as mitigations.

Hermes agent + Firecrawl

Running a local Hermes agent on Discord with Firecrawl to crawl 400 AWS doc pages in 5 minutes; migrating the agent from EC2 to a local desktop via SSH/tar orchestrated by Claude.

Local memory systems: mem0, hindsight, PG0

Embedding models and re-rankers on a local GPU, PG0 (embedded Postgres à la SQLite), hindsight semantic knowledge graphs, and CQ for team-level shared agent memory.

Timestamps

0:00 Intro & banter
2:00 Supply chain attacks: Red Hat GitHub breach, namesquatting with LLMs
7:00 VS Code plugin security model; PNPM disabling install hooks by default
10:00 AWS CodeArtifact: replacing Nexus, IAM/SSO integration
15:00 CodeArtifact RBAC — read-only devs, GitHub Actions OIDC for publishing
18:00 CodeArtifact limits vs JFrog/Artifactory enterprise features
21:00 Security vs pipeline speed tradeoffs; Datadog cost explosion
24:00 AI reshaping the security profession: checklists → automated tests & guardrails
28:00 Agentic workflows: dynamic workflows, auto mode, humans on the loop
33:00 Claude auto mode: model-based classifier for command approval
37:00 Auto mode stats — 0.4% benign blocked, 17% risky actions got through
39:00 Running agents in Docker sandboxes
42:00 Cowork for personal productivity (timesheets, contracts, HAR files)
46:00 Hermes agent on Discord + local Firecrawl stack
50:00 Crawling 400 AWS doc pages in 5 minutes
53:00 Local GPU, embedding models, PG0 embedded Postgres (SQLite-style)
56:00 Memory systems: mem0, hindsight, semantic knowledge graphs
58:00 CQ: team-level shared agent memory
60:00 Hermes PR review automation & release-please simulation
62:00 OpenClaw, risk assessment wrap-up