
Sign up to save your podcasts
Or


Quarter to seven on a Thursday: one pull request left in the queue, written by an agent, with twenty lines of code and a lockfile that moved by four thousand.
Mira and Cal follow a reviewer through the routine in the Safeguard CLI guide: a clean worktree, a review pass against the target branch, a dependency scan compared with main, questions asked in a read only mode, tests run with approval, and a review comment she posts herself. Then why code review of AI generated code fails (the quiet failure, the rubber stamp, missing evidence), a preliminary study of the versions coding agents choose, tiers for patch, minor and major upgrades, and permission rules in Claude Code. Along the way: AI coding agents, AI code security, secure coding with AI, least privilege, supply chain security for AI and a required CI check.
In the news, two stories from the last days of October and the first days of November: a trojanized copy of the Claude Code extension that was available in the Cursor extension marketplace on 29 October, and Check Point Research's 3 November report on a Codex jailbreak bundle that adds an ad to the agent's replies.
Day one on a new team. Sam installs one extension, signs in, and a warning shows up on a package she didn't choose. Mira and Cal walk through her first ten minutes with Safeguard in the editor and in the chat window: install the VS Code or JetBrains extension, sign in with the device flow, pick a remediation strategy, apply one fix, read the diff, and hand the pull request to a teammate. Then they connect Claude or ChatGPT to Safeguard's hosted MCP server and see what a fresh workspace can and cannot do.
In the news, three October stories about things that ran before a person looked: a malicious LiteLLM release on PyPI (JFrog, 21 Oct; LiteLLM's own account, 23 Oct), a planted AGENTS.md that led Codex to try installing a malicious package (Forever Security, 25 Oct), and an npm worm that armed GitHub repositories with startup hooks for Claude, Cursor and VS Code (JFrog, 26 Oct). Under the hood, the inline Package Firewall check for typosquatting, dependency confusion and malicious package names, why a hosted MCP server means there is no package to install, and what to ask about any AI coding agent. Theo brings one setting, and the episode ends with a try-this-today checklist for secure coding with AI.
Topics: AI coding agents, secure coding with AI, VS Code extension, JetBrains plugin, MCP server, MCP security, Claude connector, ChatGPT, Cursor, Claude Code, OpenAI Codex, LiteLLM, JFrog, Forever Security, AGENTS.md, npm worm, malicious packages, typosquatting, dependency confusion, supply chain security, code review of AI generated code.
For developers trying AI coding agents: your first ten minutes with Safeguard in the editor and in chat, from the VS Code and JetBrains extensions to the MCP install, with notes on Claude Code, Cursor and GitHub Copilot, MCP security, prompt injection, permissions and AI code security.
Chapters
00:00 Cold open
00:28 Intro
00:51 News: LiteLLM 1.107.0, a planted AGENTS.md, and the npm worm hooks
04:05 Sam's first ten minutes: editor and chat
08:07 Under the hood
10:02 Theo's one setting
10:52 Try this today
11:58 Outro
Sources
JFrog Security Research: LiteLLM 1.107.0 malicious release: https://research.jfrog.com/post/litellm-hijacked-again/
BleepingComputer: AI coding agents hijacked using AGENTS.md file in Do: https://www.bleepingcomputer.com/news/security/ai-coding-agents-hijacked-using-agentsmd-file-in-downloads-folder/
JFrog Security Research: beyond testmetest, 42 hijacked GitHub repos: https://research.jfrog.com/post/beyond-testmetest-42-hijacked-github-repos/
The Register: worm armed 42 GitHub repos by hiding a payload too big t: https://www.theregister.com/security/2026/10/26/worm-armed-42-github-repos-by-hiding-a-payload-too-big-to-fail/5305570
LiteLLM: The attack chain behind the v1.107.0 LiteLLM release: https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain
The Register: LiteLLM supply chain attack started with an exposed mast: https://www.theregister.com/security/2026/10/26/litellm-supply-chain-attack-started-with-an-exposed-master-key/5305544
IDE Support (docs): https://docs.safeguard.sh/docs/ide-support?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep02
MCP Integrations (docs): https://docs.safeguard.sh/docs/mcp?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep02
Topics: AI coding agents, Safeguard in the editor, VS Code, JetBrains, MCP, MCP security, prompt injection, LiteLLM, npm worm, AGENTS.md, Claude Code, Cursor, GitHub Copilot, Codex, AI code security, DevSecOps.
Ten to three on a Tuesday. Dana has one critical finding in her repo and a code freeze at five. Mira and Cal walk through her whole afternoon with the Safeguard CLI: sign in, scan on her own machine, ask the assistant for a fix in ask mode, read the prompt, review the change, open the pull request, and let a teammate approve it. Three approvals, three different questions.
In the news, two October stories: a Cloud Security Alliance note on how coding tools launch MCP servers (opencode, and GitHub Copilot in Visual Studio), and Anthropic's filing with the SEC. Under the hood, what stays on your machine, where else Safeguard shows up, and what to ask about any agent's approval prompts, with fair notes on Claude Code and Cursor. Theo brings one setting to check, and the episode ends with a try-this-today checklist.
For developers trying AI coding agents: a hands-on look at coding with the Safeguard CLI, from a local scan to a reviewed fix pull request, with notes on Claude Code, Cursor and GitHub Copilot, MCP, permissions and approval settings, AI code security and DevSecOps.
Chapters
00:00 Cold open
00:25 Intro
00:50 News: MCP launch paths and the Anthropic S-1
02:49 Dana's afternoon: three approvals with the Safeguard CLI
07:14 Under the hood, and what to ask any agent
09:17 Theo's one setting
10:17 Try this today
11:26 Outro
Sources
Cloud Security Alliance research note: MCP launch paths as RCE (openco: https://labs.cloudsecurityalliance.org/research/csa-research-note-mcp-launch-path-rce-opencode-copilot-20261/
Microsoft Security Update Guide: CVE-2026-103572: https://msrc.microsoft.com/update-guide/vulnerability/CVE-2026-103572
SEC EDGAR filing index, Anthropic, PBC Form S-1 (accession 0001628280-: https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/0001628280-26-067168-index.htm
Safeguard CLI: overview (docs): https://docs.safeguard.sh/docs/cli?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep01
Safeguard CLI: usage guide (docs): https://docs.safeguard.sh/docs/cli/usage?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep01
Safeguard CLI: interactive terminal (docs): https://docs.safeguard.sh/docs/cli/interactive?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep01
Safeguard CLI: slash commands (docs): https://docs.safeguard.sh/docs/cli/slash-commands?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep01
Safeguard CLI site: https://cli.safeguard.sh/?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep01
AI Remediate (docs): https://docs.safeguard.sh/docs/esscm/auto-fix?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep01
IDE Support (docs): https://docs.safeguard.sh/docs/ide-support?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep01
Desktop Application (docs): https://docs.safeguard.sh/docs/desktop-app?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep01
Safeguard downloads: https://safeguard.sh/download?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep01
Connect Claude to Safeguard (docs): https://docs.safeguard.sh/docs/mcp/claude?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=ep01
Topics: AI coding agents, Safeguard CLI, AI code security, fix pull request, MCP security, agent permissions, Claude Code, Cursor, GitHub Copilot, vulnerability scanning, SBOM, DevSecOps, vibe coding.
Trailer for Permission Prompt, a weekly podcast on coding with AI agents and Safeguard: hands-on workflows, tips and fair comparisons with Claude Code, Cursor, GitHub Copilot and Codex, plus MCP, IDE extensions, the CLI and the security built in, from fix pull requests to package checks. Mira and Cal work through it so you can try it today. New episodes every Wednesday.
Learn more at safeguard.sh: https://safeguard.sh/?utm_source=podcast&utm_medium=audio&utm_campaign=permission-prompt&utm_content=trailer
Topics: AI coding agents, Claude Code, Cursor, GitHub Copilot, Codex, MCP security, prompt injection, vibe coding, AI code security, DevSecOps.
From the publisher's feed