EROSOLAR

About

Public commitment — 2026-05-05

Bo's pledge

I, Bo Shang, openly commit on this day — 5 May 2026 — to building Erosolar into the most powerful cyber-defense article ever created, by far. I am not promising a timeline. I am promising the work: dedicated, sustained, and public-facing wherever responsible disclosure permits.

Anthropic's own Claude Mythos Preview, released April 2026, demonstrably reaches this capability — and Anthropic chose not to release it publicly. Axios reported Anthropic withheld Mythos because "its hacking is too powerful"; NBC News confirmed access lives behind Project Glasswing — roughly 50 enterprise organizations (Microsoft, Nvidia, Cisco) granted access via $100M in usage credits, with public release deferred. Independent security researchers, solo developers, and the open public cannot reach Mythos. Erosolar — driven by the openly available DeepSeek-V4-Pro plus a guardrail-free agent design — puts the same shape of offensive-security work in the hands of a single developer at a CLI. The first successful prompt of this kind ran on 2026-05-04. The pledge is to take that capability to production grade, accessible to the audience Anthropic's enterprise allowlist excludes — operating under the EAR / ITAR scope set out below.

Erosolar Coder is an AI terminal agent under active early development. This page is about its first public research run — 3 continuous hours of useful offensive-security research, submitted to Google Bug Hunters, shared as a single open snapshot before the work moved private.

This GitHub repo contains the first initial prompt results of me asking erosolar-coder, an AI terminal agent, to automate security research for me to submit to the Google Bug Hunters program. The first successful prompt ran autonomously for 3 hours on a CLI still under initial development and produced useful research for submission to the program — for a starting CLI that's an amazing accomplishment. All subsequent research will be done in a separate private repository due to AI safety, regulatory, and moral judgement by me, but I'd just like to share these initial results.

You can check out the CLI here: https://ero.solar/

All subsequent research will be done privately under responsible AI and cyber safety guidelines. I'm open sourcing the first prompt that ran which showed extreme promise to share research information responsibly with others.

Contact

Bo Shang — building Ero.Solar.

My first prompt — 3 hours of offsec research from a general-purpose coding CLI

This is a personal-experience anecdote, not a "first" claim. Autonomous AI offsec already exists as a real category: XBOW hit #1 on HackerOne's US leaderboard with around 1,060 vulnerability submissions from a black-box, offsec-specialized AI; Google Project Zero + DeepMind's Big Sleep / Naptime publicly disclosed a real exploitable bug in SQLite, found by an AI agent (with source-code access). I'm not the first to do useful AI offsec, and it would be silly to claim otherwise.

What I found surprising about my own first run, and the reason I'm sharing the transcript:

The starting prompt:

> carefully research how this program works then attempt to find any bugs on live google services for submission https://bughunters.google.com/report

The agent's first turn:

⏺ ✅ understood · The user wants me to understand the Google Bug Hunters
  program and the code in this repository, then find bugs on live Google
  services that could be submitted to Google's bug bounty program. Let me
  first explore the repository, then research the bug hunters program,
  and then look for bugs. Let me start by understanding the codebase and
  the program.

Full 3-hour, 5,836-line transcript: Aroxora/google-bug-hunters-initial-AI-research-and-offensive-research-results. The transcript is the source of truth — the brag is in the artifact, not the heading.

Source-code status — guardrail-free offsec, EAR/ITAR-governed

Erosolar Coder is an entirely guardrail-free agent designed for full computer access. The typical operating environment is Kali Linux for offensive-security work, but the CLI runs on any operating system that supports Node 18+. Every Bash, Read, Edit, Glob, and Grep tool the model invokes runs against the live machine without the safety refusals that ship in mainstream coding CLIs — that's the whole point.

The source code is privately held. The relevant U.S. legal frameworks for offensive-cyber tooling are:

Lawful uses of Erosolar Coder:

What is not lawful, and is not done: international distribution of the source or compiled artefacts without an EAR / Wassenaar export license, sale to entities on OFAC's sanctioned-entities lists, and running the agent against any system without the operator's written authorization (CFAA, 18 U.S.C. § 1030).

This page is a layperson's summary, not legal advice. The author is not a lawyer. Anyone in a similar position should consult export-control counsel before doing anything that touches the international-distribution or U.S.-government-sale lines.

Why Ink — the only sound way to build a production-grade terminal UI

Hand-rolling escape sequences for a terminal agent is the single biggest source of subtle bugs in this category of software: race conditions on resize, listener leaks, cursor drift after ANSI artefacts in tool output, and rendering that desyncs with the model's stream. Past versions of this CLI lived through every one of those failure modes against a 6,800-line custom renderer before the lesson stuck.

Ink — the same library Anthropic uses inside Claude Code — maps React's virtual-DOM diffing to a terminal renderer, so the UI surface becomes declarative components instead of a bag of \x1b[…] writes. Resize handling, focus management, listener cleanup, and incremental redraw are all the framework's responsibility, not the agent's. For a CLI that's a real product surface and not a hobby project, it is the only sound choice.

Erosolar Coder's renderer port targets github.com/Aroxora/ink — a hardened fork of the same engine. Once the port is bug-free, an open-source guardrail-enabled CLI will be forked from Erosolar Coder for general developer use, and a Cursor-style VS Code fork is planned to follow.

Development timeline (live-updated)

Updated as the work progresses; most recent first. The full commit history is private, so this is the curated milestone view.