Reasonable Application Security

The approval saved the password.

Hey there,

This week’s issue includes one thought on what “don’t ask again” writes to disk, five security articles worth your time, and Podcast Corner.

Chris’s Corner: the approval saved the password

What did your coding agent write down the last time someone told it to stop asking?

Say a developer on your team asks Claude Code to check something in the staging database. It proposes a command with the connection password right there in the string. They pick “Yes, and don’t ask again,” because they’ll need that query twenty more times today.

That choice saves an allow rule to .claude/settings.local.json in the project. When the rule is the exact command, the password goes in with it, in plain text.

Apiiro went looking on public GitHub. Per The Hacker News, they found 104,185 copies of that file, and 1,122 of them held at least one secret. That’s 1,379 secrets, including 379 database passwords and 23 AWS keys. Another 734 secrets were sitting only in git history.

Claude Code does try to keep the file out of git by adding it to that machine’s global git excludes. That protection lives on one laptop. It doesn’t travel with the repo, and it does nothing for a file that’s already committed. In 539 of the 1,105 repos Apiiro traced, a commit with Claude listed as co-author changed the file.

LATENT worked through Apiiro’s data and found that Gitleaks and TruffleHog together missed 19 percent of the files with secrets. A password inside a command string doesn’t look like a key.

The approval prompt is supposed to be the safety control. One click turned it into a credential store that ships with the code. Not just a credential store, but a credential treasure chest.

So this week I’d check every repo where people use Claude Code. Run git ls-files .claude to see if the file is tracked, and git log --all -- .claude/settings.local.json to see if it ever was. Then open the file and read it.

If you find a secret, rotate it first. Deleting the file doesn’t pull it back out of history or forks. Then add the file to the project’s .gitignore, and remove passwords from the commands your agent runs.

Every “don’t ask again” is a line of config. Read it like one.

Five Security Articles Worth Your Time

  1. Coinbase for Agents introduces new Agent Controls. Coinbase added limits for agents trading through its MCP: a maximum trade size, a 24-hour volume cap, and an auto-stop when a portfolio drops past a set loss. The limits run on Coinbase’s servers and can only be changed from your logged-in account with 2FA, so the agent can’t touch them. Chris’s take: New paradigm here: guardrails enforced outside of the agent’s control. I like the direction this is heading.

  2. JPMorganChase CISO Pat Opet on taking services off the open internet. On Nov 7, Opet told security teams to evaluate how to abstract their company’s services from the Internet, asking “why allow an unauthenticated anonymous agent of an attacker to interact with your systems?” In replies, he said he means authentication and authorization at the perimeter, with anonymous traffic dropped before it reaches the app, and not closing services or blocking customers’ agents. Chris’s take: Agents make probing your public endpoints cheap. I’d start with the internal tools and admin panels that never needed to face the internet.

  3. Threat actors use agentic AI tooling to find and abuse exposed secrets. Microsoft says attackers use the ARTEX agent framework to hunt public GitHub for secrets. In one case, an agent went from scanning GitHub to pulling data out of Azure Key Vault in five minutes. Chris’s take: If you aren’t aware of ARTEX today, it’s worth doing a deep dive.

  4. Hijacked Optimizely Scripts Served Card Skimmer and iOS Exploit Kits to CNN, BBC and Zoom Visitors. An attacker used a flaw in Optimizely Web Experimentation to plant a loader in customers’ Project JavaScript, served from Optimizely’s own CDN. Report URI found it in snippets for 178 organizations, swapping crypto wallet addresses and skimming cards. Chris’s take: Every third-party script runs with your page’s trust. Keep card entry in a cross-origin payment iframe, which this skimmer couldn’t read.

  5. GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold. Google found an exposed server running exploit_agent, a multi-agent framework that diffs web app code before and after patches, drafts exploits, and runs them with no human in the loop, aimed at French media and education sites. A suspected North Korean cluster also edited .claude/settings.json in victims’ internal repos, so starting Claude Code pulled down a backdoor. Chris’s take: Same file as this week’s Corner, different attacker. Review changes to agent config files the way you review changes to CI.

Podcast Corner

I love making podcasts. Here are two recent conversations.

Application Security Podcast

The Cringe Test: Why AI Can’t Do Privacy Threat Modeling for You

Dr. Kim Wuyts and Avi Douglen join Robert and me to introduce Context & Cringe, their card game for the gut-reaction privacy work it needs, and to preview APVS, privacy requirements modeled on OWASP ASVS.

“We Don’t Deal With Personal Data” (Yes, You Do)
From this episode
“We Don’t Deal With Personal Data” (Yes, You Do)

AI Security Table

Stop Patching Bugs, Just Rewrite It… Says the Guy Selling the Tokens

Cisco’s president gave a 27-year-old engineer unlimited AI tokens and 24 months to rewrite 98 percent of its legacy code. Izar, Matt, and I argue over whether disposable software is where this goes, and Izar points out who’s selling the tokens.

Not Enough Life for an AI Agent
From this episode
Not Enough Life for an AI Agent

ONE QUESTION FOR YOU

What’s in your agents’ allow lists?

Reply and tell me whether you’ve looked at the rules your developers’ coding agents saved, and what turned up.

Chris