skill
solana-incident-response-skill
Live security incident-response & war-room playbook skill for Solana (Solana AI Kit).
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/mihailShumilov/solana-incident-response-skill itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 23, 2026Solana Incident Response Skill
A Claude Code / Codex skill that turns a panicking founder or on-call engineer into an organized incident commander the moment a Solana protocol is exploited, a key is compromised, or a treasury is drained.
Built for the Solana AI Kit. Matches the kit's skill shape (progressive
skill/SKILL.mdhub +agents/+commands/+rules/) and is ready to be submoduled into.claude/skills/ext/.
⚠️ Operational support — not legal advice, not a guarantee of recovery. Every fund-moving or program-changing action is gated on a named human authority-holder. See DISCLAIMER.md.
The problem it solves
The Solana AI Kit already has deep coverage for preventing and auditing vulnerabilities — Trail of Bits, safe-solana-builder, Ghost Security, QEDGen formal verification. Nothing in the kit covers the live crisis — the minutes after an exploit when funds are leaving and a team has to decide, correctly and fast, what to pause, what to replay, who to call, and what to say.
That gap is expensive. Documented on-chain security incidents have run ~600–760 per year for three years running (~2/day); over $3.1B was stolen in H1 2025 alone; SEAL ran 125+ war rooms that year. In April 2026 the ~$286M Drift exploit drove the Solana Foundation to launch STRIDE and the Solana Incident Response Network (SIRN) — and notably, Drift's contracts were never broken: the attacker used social engineering to obtain pre-signed multisig transactions (via durable nonces) valid for over a week. Incident response is now a first-class discipline. This skill makes it a repeatable runbook instead of improvisation.
It also covers a surface the rest of the kit ignores: autonomous AI agents, which hold keys, run unattended, and fail in their own ways — from prompt injection (an agent talked into releasing funds, cf. Freysa) to silent liveness halts. See skill/agent-incidents.md.
What it does
Routes the responder through five phases plus two cross-cutting tracks, loading only the file it needs (token-efficient progressive disclosure):
| Phase | File | What it covers |
|---|---|---|
| DETECT | skill/detect.md | Confirm it's real, scope blast radius, measure velocity, snapshot evidence, start live monitoring (Helius webhooks) |
| CONTAIN | skill/contain.md | Pause/guardian, freeze authority, upgrade-authority rotation, multisig lockdown, pre-signed-tx invalidation, neutralizing upgrade — reversible-first, authority-gated |
| INVESTIGATE | skill/investigate.md | Root cause on a Surfpool mainnet fork — replay the exploit, localize the bug, verify the fix; distinguish code-bug vs key-compromise |
| TRACE | skill/trace-funds.md | Map flows (Helius Enhanced Tx), label hops (Arkham/MistTrack/Range), build the freeze-request package, hit the 72h window |
| RECOVER | skill/recover.md | Whitehat/Safe Harbor returns, patched redeploy, treasury reseed, reimbursement, resume-vs-stay-down |
| COORDINATE | skill/coordinate.md | SEAL 911, SIRN, Safe Harbor rules of engagement, exchanges, counsel, attacker negotiation |
| DISCLOSE | skill/disclose.md | Holding statement → updates → full post-mortem, timed so it never re-enables the exploit |
| PREPARE | skill/preparedness.md | Pre-incident runbook, authority roster, monitoring, Safe Harbor adoption, tabletop drills |
| AGENT INCIDENTS | skill/agent-incidents.md | Autonomous-agent failure modes (prompt injection, key compromise, liveness halt, MEV, oracle, contagion), deterministic trigger taxonomy, pre-deploy risk scoring, risk-transfer options |
Agents
| Agent | Model | Role |
|---|---|---|
incident-commander | opus | Sequences the phases, gates irreversible actions on humans, keeps the war-room log |
forensics-analyst | opus | Surfpool replay + root cause + fund tracing |
comms-officer | sonnet | Holding statements, updates, post-mortem |
Commands
| Command | Purpose |
|---|---|
/ir-triage <tx | symptom> | Classify the incident + first-15-minutes action list |
/ir-contain | Authority-gated containment plan with exact txs |
/ir-trace <address> | Fund-flow map + freeze-request package |
/ir-postmortem | Draft the post-mortem from the war-room log |
/ir-drill | Tabletop exercise to pressure-test the runbook |
Templates
war-room-roster.md (append-only log) · post-mortem.md · user-comms.md · safe-harbor-checklist.md
Safety model
This skill is read-only by default: it investigates and proposes, and never signs, sends, freezes, or upgrades on its own. Every irreversible action is drafted as an exact transaction routed to a named authority-holder to approve and execute. It preserves evidence before changing state, reproduces fixes on a fork before mainnet, and never contacts an attacker without an agreed strategy. Full rails in rules/incident-response.md.
Why it fits the kit (no duplication)
It consumes the kit instead of repeating it: vulnerability classes/fixes → ext/trailofbits, ext/safe-solana-builder, ext/qedgen; forked replay → the Surfpool MCP + /debug-user-tx; on-chain data → the Helius MCP; legal follow-through → crypto-legal-skill. This skill is the playbook and coordination layer that orchestrates them during a live crisis — the missing piece.
Installation
git clone https://github.com/<you>/solana-incident-response-skill
cd solana-incident-response-skill
./install.sh # interactive, sensible defaults
# or
./install.sh -y # non-interactive
Installs to ~/.claude/skills/solana-incident-response/ (mirrors to ~/.codex/skills/ if the Codex CLI is detected). Use ./install-custom.sh to choose personal vs. project scope, or set CLAUDE_SKILLS_HOME=/path ./install.sh. No network calls, no build step, no dependencies.
Use in Claude Code
After install, just describe the situation ("our vault is draining", "an unexpected upgrade landed") and the skill activates, or invoke a command directly: /ir-triage <tx>.
Calendar versioning
Security context moves fast, so releases are calendar-pinned, not semver. Current: 2026-06. Last source review: 2026-06-21. See resources.md for the dated reference set.
Contributing
PRs welcome — especially new containment primitives, updated responder contacts, fresh reference incidents, and counsel review of the coordination/disclosure guidance. Keep claims diffed against primary sources.
License
MIT — see LICENSE. Built for the Solana AI Kit community challenge.