Agents that act. Rails that hold.
I'm Drew, a cyber architect. I build autonomous systems that take real actions in production and the guardrails that make that a safe bet.
Founder at DCYFR Labs · Head of AI at GameShark Labs · Principal Security Engineer at Monks
cyberdrew.dev is the long version of this page.
60+ agents on one substrate · continuous autonomous cycle · local-first default model tier · 6+ yrs security engineering
@dcyfr/aiopen source
The portable TypeScript runtime the fleet thinks in: model routing across local and frontier tiers, tool-use, and MCP wiring you drop into an existing project.- The agentic fleet
running
Sixty-odd daemons on a shared substrate, spanning research, code review, monitoring, and self-healing. They run autonomously on a local-first model stack under a hard budget. cache-invariantsopen source
Finds the code that silently breaks LLM prompt caching: a linter for agent code and an audit that replays real call logs.spend-gateopen source
A budget trip for autonomous agents. When spend crosses a threshold, every paid call that checks the gate stops, and the trip survives a Redis outage or restart.- SharkVault
live
Proof the fleet ships product and not just diffs: a backer-funded consumer app taken end to end, from auth to payments to content pipeline, largely by agents.
Starter templates for agents, APIs, GraphQL, React, chatbots, CLIs, and Node.js live in the DCYFR Labs org, which holds the canonical repo catalog.
The fleet doesn't wait to be asked. It wakes on a schedule, reads its own state, picks work off the queue, and writes down what it learned. Then it starts over. Five states, each with something that can stop it.
| State | Runs on | What stops it |
|---|---|---|
| 01 Observe | no model | Restricted paths never enter context: message stores, journals, backups |
| 02 Triage | tier-0, local | Local-first routing. Frontier tiers cost money, so they have to be earned |
| 03 Think | tier-3, frontier | A metacognition breaker halts any agent spinning past six unproductive cycles |
| 04 Act | sandboxed | Allowlisted binaries only. No metacharacters, no pipe to rm, no private-IP fetches |
| 05 Record | append-only | Writes confined to two directories. Everything else is read-only to the daemon |
Autonomy is a claim about what happens when nobody is watching. These are the controls that make the claim checkable.
| Control | What it means |
|---|---|
Spend ceiling (enforced) |
A hard monthly ceiling on model spend, sized to the deployment. Trip it and the kill switch fires. The pattern is published as spend-gate. |
Sender kill switch (< 1 s) |
Every outbound path (mail, chat, webhooks) honours one pause flag. |
Credential isolation (per-process) |
Secrets resolve at exec time from the OS keychain. Nothing lands in an environment dump, a plist, or a prompt. |
Tool sandbox (allowlist) |
The shell an agent gets is not your shell: fixed binaries, no metacharacters, no SSRF to localhost or private ranges. |
Metacognition breaker (6 cycles) |
An agent that keeps picking the same task without shipping is quarantined. |
Self-heal watchdog (3 / hr) |
Downed services restart on graduated trust. Dry-run first, and never more than three times an hour before escalation. |
When something I depend on is broken, the fix goes back to the project. This list is rebuilt weekly from GitHub search by scripts/upstream-ledger.mjs. It shows merged and open pull requests only.
2 merged and 6 open across 6 projects in the last 12 months.
| Project | Pull request | Status |
|---|---|---|
| batocera-linux/batocera.linux | #16444 rk3568: restore the Rockchip GMAC (stmmac) driver in the kernel config | merged 2026-09-12 |
| batocera-linux/batocera.linux | #16445 rk3568: fix ODROID-M1S Ethernet by naming the RTL8211F by PHY ID | merged 2026-09-12 |
| Clownacy/clown68000 | #5 Give ADDQ and SUBQ a defined source address mode | open since 2026-10-08 |
| Clownacy/clownmdemu-libretro | #2 Export only the libretro API from the unix build | open since 2026-10-08 |
| libretro/supafaust | #35 libretro: expose work RAM as RETRO_MEMORY_SYSTEM_RAM | open since 2026-10-08 |
| libretro/supafaust | #36 libretro: accept every cheat format, and apply RAM-write cheats | open since 2026-10-08 |
| anthropics/claude-code-security-review | #128 fix: use configured model in validate_api_access instead of retired claude-3-5-haiku | open since 2026-08-29 |
| drcursor/HushScribe | #5 Surface system-audio capture failures instead of swallowing them | open since 2026-08-27 |
Notes from the build live at dcyfr.ai/blog: agentic AI security, autonomous delivery, and what breaks when you let agents act. There is a feed if you would rather subscribe than visit.
GIAC · GDSA Defensible Security Architecture · GSTRT Strategic Planning, Policy & Leadership · GCIH Certified Incident Handler
CompTIA · SecurityX Advanced Security Practitioner (CASP+)
GitHub Sponsors pays for the open-source work on this page and the upstream fixes above. Tiers start at $5 a month, and every sponsor gets a monthly build note: what shipped, what broke, what's next.
I work with teams putting agents into production. That usually looks like one of three things: your first secure agent, a governed fleet, or a hard look at what you already run.
I'm not taking paid engagements right now, and advice is pro bono.
Book an intro call · hello@cyberdrew.dev · LinkedIn · X
For security disclosures: security@dcyfr.ai (see the DCYFR Labs security policy).



