LLM-native vulnerability discovery as a Superpowers fork.
Redpowers is a fork of obra/superpowers that adds four skills for LLM-based vulnerability discovery. It is inspired by the Vidoc Security Lab article "We Reproduced Anthropic's Mythos Findings With Public Models" (April 2026), which showed that Mythos-style vulnerability-discovery workflows are reproducible with public frontier models when the orchestration is right.
Redpowers implements a Planning → Detection → Validation → Triage pipeline as four Superpowers-compatible skills. It is purely LLM-driven — no Semgrep, no CodeQL, no external SAST. The workflow relies on subagent dispatch, hypothesis-driven detection, and multi-dimensional validation (root cause, data-flow reachability, impact fit, realistic trigger, false-positive check).
Focus: Web/SaaS code in Node, Python, and Go. Mobile, native, IaC, and CI/CD are out of scope for v0.1.0.
Redpowers builds on obra/superpowers by Jesse Vincent. All Superpowers skills (brainstorming, writing-plans, executing-plans, subagent-driven-development, test-driven-development, systematic-debugging, verification-before-completion, dispatching-parallel-agents, requesting-code-review, receiving-code-review, using-git-worktrees, finishing-a-development-branch, writing-skills, using-superpowers) are preserved unchanged. The four new security-review-* skills are fork-specific additions.
Upstream receives none of these changes. The upstream project explicitly rejects PRs that "rebrand the project, add fork-specific features, or merge fork branches" — see CLAUDE.md in this repo for the full upstream contribution policy (preserved verbatim from Superpowers). If you want the upstream core, install obra/superpowers instead. If you want the security-review additions, install Redpowers.
Attribution: the Superpowers engineering philosophy, skill conventions, and the skill-discovery mechanism are all Jesse Vincent's work.
Note: Installation differs by platform. Substitute RCXDd/redpowers for obra/superpowers in any upstream install instruction you already know.
Register the Redpowers marketplace, then install:
/plugin marketplace add RCXDd/redpowers
/plugin install redpowers@redpowers-devClone and symlink (mirrors the upstream Codex install):
git clone https://github.com/RCXDd/redpowers.git ~/.codex/redpowers
mkdir -p ~/.agents/skills
ln -s ~/.codex/redpowers/skills ~/.agents/skills/redpowersWindows (PowerShell):
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.agents\skills"
cmd /c mklink /J "$env:USERPROFILE\.agents\skills\redpowers" "$env:USERPROFILE\.codex\redpowers\skills"Restart Codex to discover the skills. See .codex/INSTALL.md for details (the file still references superpowers in places — treat it as a template).
/add-plugin redpowers
Or search for "redpowers" in the Cursor plugin marketplace.
Add Redpowers to the plugin array in your opencode.json:
{
"plugin": ["redpowers@git+https://github.com/RCXDd/redpowers.git"]
}Restart OpenCode. See .opencode/INSTALL.md for migration notes.
copilot plugin marketplace add RCXDd/redpowers
copilot plugin install redpowers@redpowers-devgemini extensions install https://github.com/RCXDd/redpowersIf you also have upstream Superpowers installed, the two plugins coexist — Redpowers adds four new skill directories and renames the plugin namespace from superpowers to redpowers.
| Skill | What it does |
|---|---|
security-review-threat-model |
Once-per-repo attack-surface mapping. Produces .security/threat-model.yaml covering entry points, trust boundaries, sensitive data, external dependencies, and a ranked priority_files list. |
security-review-file |
Three-pass per-file detection: Planning (semantic chunking by attack surface) → Detection (parallel subagent per high-priority chunk, with cross-file read access) → Validation (five-dimension check: root cause, reachable data-flow, impact fit, realistic trigger, false-positive). Emits findings against skills/security-review-file/schemas/finding.schema.yaml. |
security-review-triage |
Consumes raw findings. Deduplicates by root cause, detects exploit chains (e.g., authz-bypass + SSRF), and produces a P0-P3 Markdown report ranked by reachability rather than CVSS. Includes chain-elevation: two P2 findings that chain can escalate to P1. |
security-review-repo |
Whole-repo orchestrator. Requires .security/threat-model.yaml to exist (runs security-review-threat-model first if missing), reads priority_files, spawns security-review-file per file, and honours a cost handshake before expensive multi-file runs (default top-20). |
Each skill has its own SKILL.md under skills/<skill-name>/ with a frontmatter description field that encodes the trigger phrases.
skills/security-review-patterns/ is reference documentation, not a skill. It contains Markdown files organised by bug class (currently web-auth/missing-authz.md and web-injection/sql-injection.md, with more added iteratively). These are loaded on-demand by detection subagents when a finding hypothesis matches a known pattern — progressive disclosure, not auto-preloaded.
The directory has no SKILL.md and is intentionally not registered as a skill. If your harness auto-discovers anything under skills/ as a skill, check that it keys on SKILL.md presence (all mainstream harnesses do); if not, add security-review-patterns to your ignore list.
The skills are trigger-driven. Example prompts that activate each skill:
- "Build a threat model for this repo" →
security-review-threat-modelruns, produces.security/threat-model.yaml. - "Audit
src/handlers/orders.tsfor security issues" →security-review-fileruns its three passes and emits findings. - "Security scan this codebase" or "Find all vulnerabilities in this repo" →
security-review-repokicks off the orchestrator (with cost handshake). - "Triage these findings" →
security-review-triageconsumes the finding YAMLs and produces the P0-P3 report.
Triggering is description-driven — the skill frontmatter description lists the canonical triggers. Phrasing the request with the right verb ("audit", "threat model", "triage", "security scan") is the reliable path.
From PLAN.md §12:
- No exploit generation. Redpowers discovers and validates; it does not produce weaponised PoCs.
- No auto-fixes. Remediation suggestions are advisory prose; Redpowers does not patch code (that's CodeMender territory).
- No external security tools. No Semgrep, CodeQL, Snyk, etc. LLM-only.
- No crypto-primitive strength evaluation. Redpowers flags misuse patterns (weak hash, hardcoded secrets); it does not cryptanalyse.
- No IaC checks. Terraform / CloudFormation / Kubernetes manifests are out of scope.
- No mobile or native code. Web/SaaS only for v0.1.0.
- No CI/CD hook. Redpowers runs inside a coding-agent conversation, not as a Git hook or CI step.
Superpowers = Engineering (correctness, TDD, green-first). Redpowers = Adversarial Engineering (violating assumptions, red-first, edge-case paranoia). Some Superpowers principles invert under red-team pressure; see docs/philosophy.md in the repo.
The Superpowers engineering philosophy is otherwise unchanged:
- Test-Driven Development — skills are developed via pressure-tests (adversarial prompt-engineering eval)
- Systematic over ad-hoc — process over guessing
- Evidence over claims — verify before declaring a finding valid
- Progressive Disclosure — skill metadata preloaded, detail loaded on demand
Redpowers does not track upstream Superpowers version numbers. This is v0.1.0 — first fork-specific release. Upstream may be at v5.x; Redpowers uses its own semver starting from 0.1.0.
To install a pinned Redpowers version in OpenCode:
{
"plugin": ["redpowers@git+https://github.com/RCXDd/redpowers.git#v0.1.0"]
}Contributions to fork-specific files (skills/security-review-*, skills/security-review-patterns/, tests/pressure-tests/, fixtures, and this README) are welcome against github.com/RCXDd/redpowers.
Do not send Redpowers changes upstream to obra/superpowers. The upstream project has a 94% PR rejection rate and explicitly rejects fork-specific features and rebranding. See CLAUDE.md in this repo for the full upstream contribution policy — it is preserved verbatim and applies only to upstream PRs, not to fork contributions.
If you want to propose a new security-review pattern, add a Markdown file under skills/security-review-patterns/<category>/<slug>.md and open a PR against this fork.
MIT License — same as upstream Superpowers. See LICENSE.
- Vidoc Security Lab — "We Reproduced Anthropic's Mythos Findings With Public Models" (April 2026)
- Anthropic Frontier Red Team — "Assessing Claude Mythos Preview's cybersecurity capabilities"
- Anthropic — "Project Glasswing"
- OpenAI — Aardvark / Codex Security (March 2026 preview)
- Google — CodeMender
- OWASP Application Security Verification Standard (ASVS)
- GitHub Advisories Database — source of smoke-test fixtures
- obra/superpowers — upstream project
- Jesse Vincent — original Superpowers release announcement
Redpowers would not exist without Jesse Vincent's Superpowers framework. The skill-discovery mechanism, the subagent-driven-development pattern, the pressure-test methodology, and the voice (including "your human partner") are his. Redpowers adds four skills and a pattern library; the scaffolding underneath is Superpowers.
Redpowers is not affiliated with or endorsed by Anthropic, Vidoc Security Lab, OpenAI, Google, or Prime Radiant.