[HackerNotes Ep. 165] Protobuf Hacking, AI-Powered Bug Hunting, and Self-Improving Claude Workflows

Hacker TL;DR

• Protobuf parameters on Google are an underexplored attack surface. Claude excels at decoding and re-encoding them

• Login CSRF chained with browser permissions (mic/camera) can escalate to full surveillance of victims

• Add a self-improving loop to your Claude.md file so it learns from mistakes and saves tokens over time

• Models now score ~69% on CyberBench. AI-assisted hacking is no longer hype; it's producing real bounties

  

Today's Sponsor: Check out ThreatLocker Ringfencing https://www.criticalthinkingpodcast.io/tl-rf

In the News

• BB Scope Major Update - @sw33tLie shipped a big upgrade to BB Scope, now with the ability to dump your HackerOne reports via a new pull request, useful for feeding context to AI tools

• Caido Skills for Claude Code - Caido officially released a Claude Code skill with SDK integration for searching HTTP history, adding findings, editing requests, and managing match-and-replace rules

• Matt Brown Goes Full-Time - @nmatt0 left his job to go full-time into hardware hacking, pentesting, and content creation. His YouTube channel is the go-to resource for IoT device hacking

Protobuf Hacking on Google

Justin has been finding consistent success targeting protobuf-encoded parameters across Google's attack surface. Many parameters are base64-encoded binary protobuf structures that most hunters skip, either assuming they're signed or simply not wanting to deal with the complexity.

How Protobuf Encoding Works

Each field in a protobuf message is encoded with a single byte containing three pieces of information:

• Bits 0-2 (Wire Type): Defines the data type like varint, length-delimited string, fixed-size integer

• Bits 3-6 (Field Number): The index of this field in the protobuf schema

• Bit 7 (Continuation Flag): Indicates whether the metadata extends into the next byte

Without the .proto schema definition, you won't get field names, only field numbers and values. But by decoding nested protobuf structures, you can identify IDs, strings visible in the UI, and internal references that unlock new attack vectors.

So you can use Claude to decode protobuf blobs. It handles nested structures, identifies checksums, and can generate Python scripts to re-encode modified payloads. Justin built a Caido convert workflow to do inline protobuf decoding, which he'll share on the Discord.

We do subs at $25, $10, and $5, premium subscribers get access to:

– Hackalongs: live bug bounty hacking on real programs, VODs available
– Live data streams, exploits, tools, scripts & un-redacted bug reports

You can also find some hacker swag here.

Login CSRF + Browser Permissions = Surveillance Chain

Justin shared a high-impact chain he's been reporting successfully across multiple targets:

1. Login CSRF into the attacker's account on the target application

2. CSRF the audio recording endpoint to start capturing the victim's microphone

3. The application transcribes the audio and uploads it to the attacker's account

If the victim has previously granted the site microphone permissions (a domain-level browser setting), this requires zero user interaction. The attacker can silently record and exfiltrate audio.

Browser permissions are scoped to the domain, not to the account. So a login CSRF effectively hijacks those permissions.

Iframe Permission Delegation

Check the allow attribute on iframes. If a parent page delegates permissions like microphone, camera, or display-capture to an iframe you can control, you may be able to abuse those permissions from a third-party context within a trusted top-level page.

AI Hacking Debate: Skill Gaps and False Positives

Can anyone just spin up Claude Code and start finding bugs nowadays ?

What AI Unlocks

• Claude Code + xssdoctor found a high/critical on a major company by testing survey endpoints, the kind of boring functionality experienced hunters habitually skip

• BusFactor had Claude find a high on a program paying $8K for highs, one bug pays for three years of usage

• The gap between technical and semi-technical users has shrunk dramatically, but it hasn't disappeared

Where Human Expertise Still Matters

• False positive triage: When Claude says "CRITICAL FINDING," there's roughly a 20% chance it's actually valid. Knowing that a wildcard CORS isn't exploitable or that a finding is a medium rather than a critical requires real security knowledge.

• Validation complexity: Even experienced hunters need time to fully understand and validate the unusual bugs AI discovers

• Context and judgment: Setting up infrastructure, understanding scope, and knowing what's worth reporting still requires a hacker's intuition

The Self-Improving Claude.md Trick

Joseph shared a practical tip for leveling up your Claude Code setup:

Add this to your CLAUDE.md:

This creates a self-improving loop: every session makes the next one better. Claude stops wasting tokens on repeated mistakes and learns patterns specific to your workflow, like where session files are stored or which commands work on your system.

Discord Bot for Remote Hacking

Instead of using Claude RC (which doesn't support --dangerously-skip-permissions), Joseph built a Discord bot that acts as a remote Claude Code interface:

• Each new task spawns a Discord thread as a session

• Tool calls are displayed with diff-formatted code blocks (green for additions, red for removals)

• Supports voice messages and attachments for input

• Includes a --resume command at the top of each thread for jumping back into sessions from a VPS

• Acts as a DevOps engineer on speed dial — validate findings, check logs, host files, all from a phone

Resources

• BB Scope - Dump scope and reports from HackerOne/Bugcrowd with API key support

• Caido Skills - Official Claude Code skill for Caido proxy integration

• Matt Brown's YouTube - Best hardware hacking content on YouTube

• Matt Brown's Courses - Digital Signal Analysis for Hardware Hackers ($200) and Beginner's Guide to IoT Hacking ($50)

• MDN: iframe allow attribute - Documentation on permission delegation to iframes

• Anthropic Privacy Settings - Disable training on your data before feeding reports to Claude

That's it for the week, keep hacking!