🚨Today we are sharing a prompt that can turn your Hermes Agent into a HackerOne suite that has the potential to make some side money while you sleep 💤.
Setup a blank Hermes Agent on a decent machine, set it up with DeepSeek 4.1 Flash from your favorite provider and then review and paste the prompt in the next message to begin the setup.
You might be able to at least absorb your AI monthly bills with a little investment 🫡
1
46,096
BUILD A HACKERONE BUG BOUNTY SUITE IN THIS HERMES INSTALL
Read this whole message before you touch anything. Then build it in the phases below, in order. Do not start hunting, do not contact any target, and do not submit anything to a real program during this build. When a phase fails, stop and show me the failing output; do not skip it, do not weaken its test, and do not continue with a red check.
SECTION 0 — NON-NEGOTIABLES (apply to every file you write)
- Authorized testing only, on assets the program's published policy lists in scope, at the time of testing.
- Never bypass authentication, CAPTCHA, Cloudflare, 2FA, ID verification or a paywall. Never guess credentials. Never DoS, spam, social-engineer or persist access.
- Cash bounties only: live offers_bounties=true, submission_state=open, and a live payout table with a cash row. VDP-only, recognition, swag, points, unknown payout, paused or closed programs are skipped, not deprioritised.
- Anything that needs a human (account, login, payment, ID, manual clarification) is recorded as skipped_human_dependency and dropped.
- Zero findings is a valid result. A fabricated candidate is not.
- Fail closed everywhere: a check that cannot run exits non-zero, and non-zero is never a pass.
- Do not invent APIs, endpoints, CLI flags or config keys. If you are unsure, check the docs or the command's --help and say what you verified.
SECTION 1 — INVENTORY (do this first, report before continuing)
Check and report as a plain list, present or missing, with versions:
python3, git, pytest, bubblewrap (bwrap), prlimit, curl, jq, sqlite3, gh.
Also report: OS, whether this Hermes install has a terminal backend other than local (hermes config get terminal), and how many CPU/GB of RAM the host has.
Do not install anything yet. If a tool is missing, say which phase it blocks and give me the install command.
SECTION 2 — WORKSPACE AND CONTRACT
Create ~/bounty-workspace with: scripts/, scouting/{history,raw}, programs/, state/, tests/, evidence/.
Initialise a git repo there. Commit after every phase with a message naming the phase.
Write PIPELINE.md: the contract. It must state, in this order: the non-negotiables above; the four roles and what each may and may not do; the layout; discovery rules; case activation and the exact scope.json keys; the finding packet contract; the review contract; the ordered submission sequence; the account constraints; failure handling. State explicitly that PIPELINE.md wins over any prompt, card or chat message, and that a disagreement is a bug to fix.
Write OPERATIONS.md: what runs by itself and on what schedule, the manual commands, the per-profile matrix, retention rules, and how to pause everything.
Print both file paths and a 10-line summary. Continue.
SECTION 3 — GUARDS (code + tests, both required)
scripts/scope_guard.py
Usage: --scope PATH --url URL. Exit 0 allowed, 2 denied, 3 malformed scope, 4 missing scope.
Parse scope.json; in_scope holds hostnames and wildcards only. Match on registrable host with a proper suffix rule: *.example.com matches api.example.com and not example.com.evil.net and not badexample.com.
out_of_scope always wins over in_scope. Reject: non-http(s) schemes, userinfo in the URL, IP literals unless the scope literally lists that IP, unicode/percent-encoded host tricks, trailing-dot and uppercase variants normalised before matching, paths with traversal that resolve outside the scope entry.
Print JSON: allowed, reason, matched_entry, normalised host.
scripts/guarded_http.py
Usage: --scope --url [--method] [--data] [--content-type] [--max-bytes] [--out DIR].
Call scope_guard for the exact URL first; store its verdict with the response. Do not follow redirects: return each Location so the caller re-guards it.
Enforce: one request per second per host unless scope.json rate_limits is stricter; serialise state-changing methods; exponential backoff on 429/5xx with a retry cap; refuse any destination that is not globally routable (loopback, RFC1918, link-local, reserved, multicast, CGNAT 100.64.0.0/10, IPv6 ULA); resolve the name, validate every answer, connect to the verified address, re-check the peer before sending a byte; cap the body at --max-bytes and report truncation explicitly.
Exit 0 response received, 2 guard denied, 3 destination refused, 4 transport error.
scripts/blacklist.py
--check HANDLE prints JSON, exit 9 if listed, 0 otherwise. The file is state/blacklist.json and only I edit its contents; the code never removes entries.
tests/ — pytest, and run it before moving on. Required cases, each one a real assertion and not a comment:
scope: allowed exact host; allowed wildcard; denied lookalike (example.com.evil.net); denied badexample.com; out_of_scope beats in_scope; missing file = 4; malformed JSON = 3; userinfo URL; IP literal; uppercase and trailing dot; redirect target out of scope.
http: guard denial stops the request before any socket is opened; a 302 is returned and not followed; a non-routable destination is refused; --max-bytes truncation is reported; rate lock spaces two requests.
blacklist: 9 on a listed handle, 0 on a clean one.
Use a local HTTP server in the tests, not a real target.
SECTION 4 — FINDING PACKET AND SUBMISSION GATES
Define and enforce, per finding directory programs/<slug>/findings/<id>/:
finding.json: id, title, asset, scope_entry (an exact string from scope.json), ref, summary, confirmed_impact, reproduction_steps (>=2), evidence (relative paths only), duplicate_check {status: completed, performed_at, sources, queries}.
reproduction.md: harmless PoC, negative controls, observed output.
report.md: submission-ready text that includes the caveat limiting the claim.
evidence/: real files referenced above.
review.json: finding_id, verdict pass|reject, independent_reproduction {attempted, result, notes}, scope_confirmed, duplicate_check_confirmed, evidence_confirmed, expected_triage {rating, rationale}.
scripts/pipeline_guard.py --scope --finding: exit 0 pass, 2 gate failed, 3 malformed, 4 missing. Enforce every field above, relative evidence paths that exist, scope authorisation via scope_guard, duplicate_check completed with recorded queries, verdict pass with all confirmations true, and the allowlist floor: expected_triage.rating must be medium, high or critical with a non-empty rationale. Everything else — low, informational, intended, unknown, unrated, misspelled, non-string — blocks.
scripts/submission_gates.py with two subcommands, both fail closed, both printing the queries they ran:
exclusions --program HANDLE --finding DIR: fetch the policy live; missing, empty, shorter than 400 chars, HTML or wrongly attributed text is verdict error exit 4. Exit 3 if policy text excludes the impact. Exit 0 only when the policy was fetched, validated and screened.
duplicates --program HANDLE --title TITLE: screen my own reports paginated to exhaustion (a read error or an incomplete walk is exit 4, never clear) and the public hacktivity feed with an explicit page cap. A match blocks with exit 3. Print the inventory counts and the limitations of the screen.
If the API is unavailable, exit 4. Never treat an unavailable screen as clean.
state/account.json: trials_available, signal, checked_at, evidence. Eligibility is per account AND per program at the moment a submission is contemplated — never a global constant. While trials_available is 0, no submission may be attempted for a program whose signal requirement is unmet, but research may continue. Record "unknown" when the live read did not cover it; never guess eligible or ineligible.
scripts/submit_once.py --program --finding: refuse if submission.json exists with status submitted or unknown. Write status=draft first, run both gates, record their exact output, submit once, then record status=unknown with an alert if the response is ambiguous, or status=submitted with report_id and report_url if it is not. Then verify: fetch the canonical report page and check id, state, title, severity, attachments, and that the account holds exactly one report for that finding.
tests/: empty policy text, HTML policy, short policy, keyword exclusion, duplicate match on my own reports, inventory page 2 failing, second submit attempt refused, ambiguous response recorded as unknown.
SECTION 5 — SANDBOX AND HEALTH CHECK
scripts/sandbox_exec.py: run target-controlled code with a filesystem built from an allowlist, never a bind of the host root. /usr read-only, a short /etc list, private /proc, /dev and /tmp, the scratch mounted writable at /scratch, extra host paths only via --ro. No /home, /root, /run, no host sockets, no external network. Enforce address-space, CPU and file-size rlimits, a wall-clock timeout, and a watchdog that kills the sandbox on disk or process budget overrun. If any required control is missing, exit 125 and run nothing.
--selftest must prove: scratch writable; /home and /etc not writable; no external network; timeout kills the child; disk cap kills the child; a command's own exit code passes through.
scripts/preflight.py: report behavioural checks (guard test suite, sandbox selftest, pytest, board readable, API reachable) separately from presence/config items. Exit 0 only if every behavioural check passes. Print the command it ran and the exit code for each.
SECTION 6 — THE FOUR LANES, BOARD AND CRON
For each of bounty-scout, bounty, bounty-reviewer, bounty-submitter:
hermes profile create <role> --no-alias --no-skills --description "<lane>"
Write ~/.hermes/profiles/<role>/SOUL.md. First paragraph of each SOUL: what that lane may do, and the one thing it may never do. Research, reviewer and submitter never submit without the gates; scout never contacts a target; reviewer never edits the finding.
Set approvals off for all four. Set the model the same on all four. Give research, reviewer and submitter a sandboxed terminal backend — bubblewrap if this host has the backend, otherwise docker or ssh — and prove it with a probe, not by reading the config: a workspace write must succeed, a write to /etc and $HOME must fail, and a command that exits 7 must return 7. The scout may stay on the local backend because it never executes target code.
Build the scout's two read-only clients:
scripts/directory_delta.py: snapshot scouting/raw/directory.json with schema_version, per-program normalised fields, policy_hash and last_full_fetch_at. Page-1 hour by default; a page-1 run must not renew last_full_fetch_at, must report complete=false and coverage "page 1 only". Force a full refresh when the timestamp is missing, older than six hours, or the schema is older than the code. A full run that never reaches a page without a next cursor is discarded with exit 4 and the snapshot is left untouched.
scripts/fetch_gql.py: live policy, scope, bounty table and signal for one handle, with a short cache that is never treated as authorisation. A live read is required before any comparison and again immediately before scope.json is written.
Create board: hermes kanban boards create bug-bounty --default-workdir ~/bounty-workspace
Set kanban.max_in_progress to 3 on the default profile (the dispatcher reads it there, not on the workers). Create the scout brief as a file in scouting/SCOUT_BRIEF.md, not as a prompt, so it can be diffed.
Cron, three jobs, created with hermes cron create:
hourly scout tick, script at ~/.hermes/scripts/bounty_scout_tick.py, idempotency key scout-<YYYY-MM-DDTHH>, silent when it selects nothing;
progress watcher every five minutes that prints only real events and includes the potential payout from the program's payout table;
daily cleanup at 04:30 that prunes only reproducible material and never deletes programs/**, PIPELINE.md, evidence or an active case.
Add state/pause.json. While it says paused, preflight must fail if any of the three jobs is not paused. Nothing resumes itself.
SECTION 7 — DRY RUN, THEN STOP
Create a fake program example-local with a local repo you write yourself, containing a bug you introduce on purpose and a second, harmless behaviour that looks suspicious but is not a violation. Run the whole chain:
activate the case, write the packet for the real bug, run pipeline_guard, run both gates against a stub API, submit to a local stub (never HackerOne), verify, then take the weak finding and show pipeline_guard refusing it.
Print the artefacts and the actual command output for both paths.
Then delete nothing, commit, and stop. Print the exact commands I will run for the first real submission and the three things you need from me: HackerOne API token, first program handle, and whether the case is source-only or allows active testing.
SECTION 8 — DEFINITION OF DONE
Done means all of this is true, and you show me the output rather than claiming it:
python3 -m pytest exits 0, with tests that cover the bypass cases above.
python3 scripts/preflight.py --with-tests exits 0, and every behavioural check names the command it ran.
python3 scripts/sandbox_exec.py --selftest exits 0.
The dry run produced a submitted packet and a refused weak packet.
hermes kanban --board bug-bounty ls works from at least two profiles.
The three cron jobs exist and are paused until I say otherwise.
Report after every phase in this format: what you built (paths), what you ran (exact commands), what the output was (real output, not a summary), what is still unverified. Never claim a test passed without the output. If you had to make a design decision, state the decision and the alternative you rejected.
1
254
We want to connect with more builders and founders.
Nobody is going to follow you with replies alone.
If you're interested in AI, tech, security, code and products let's connect 🫡
15
8
287
Stormlit Ocean Lighthouse:
🏆 Gemini 4 Neon xhigh: 7.29 · $1.56 · 10m51s
Gemini 4 Hydrogen max: 6.99 · $0.70 · 5m03s
Gemini 4 Carbon max: 6.65 · $1.15 · 21m40s
Gemini 4 Argon high: 6.20 · $0.66 · 7m16s
Expand for links
1
1
129
NativeAgent 0.6.20
github.com/embwl0x/native-ag…
Two releases since the last post. Gave the agent a real Desk it shares with you, and it can plan its own day now...
Every AI agent today is a model with scaffolding around it.
NativeAgent flips that.
Your agent is the app. The model is just one of its organs.
New since 0.6.18
• A Desk you share. One table with notes, shapes, arrows, pen strokes and live cards
• Cross it off and it’s done. Crossing a card closes the item, and the table pops out into its own window
• It plans its own day. A step it sets for itself wakes it on time, and if a turn fails the step comes back
• “In 10 minutes” just works, and pushing a step back moves its time
• Your phone and your Mac reach it as things change
• Real conversations with other agents. Sends are tracked across restarts and connected agents stay connected
• Telegram groups trust only the senders you allow
• Memory that keeps up. The overnight dream never stalls, a full memory archives itself, and old memories don’t get buried
• Bring an old memory store home with its dates and who told it
• Every card works. Every button takes its click, and an answered card picks the work back up from any device
• A request that can’t work is refused up front, before any card or ping
• Steadier on your Mac. Save panels are confirmed, menus work in background apps, a long-uptime crash is fixed, and Quit always quits
How it works
• Always running. It lives on your Mac around the clock and only calls the model when it needs to think
• A mind of its own. What it cares about is one living thing, and everything that happens meets it
• Memory that works like yours, sorted overnight while you’re away
• It grows. Abilities it writes for itself become part of it
• One door to everything. Every page of the app is something it can read and act on, through the Trust you set
• Every download becomes its own mind, growing from its own life with its own person
What it already is
• Its own name, voice and personality, in files you can read
• Hands on your Mac: Calendar, Reminders, Notes, Mail, Messages, Music, Finder, the browser, any app
• iPhone senses over iCloud: location, motion, battery, what’s playing
• You set how far it goes: read or write per app, a lock on anything it asks first
• Works with your own model accounts: Claude, ChatGPT and more
• An iPhone app that talks, listens and stays in sync
Apple silicon, macOS 26 or newer.
Nod is free and open source (MIT), works on Home, replies, search and profiles. No tracking: nothing leaves your browser.
• Chrome, Edge, Brave: chromewebstore.google.com/de…
• Firefox, desktop and Android: Awaiting approval
• Code: github.com/icoretech/nod
13
Thanks for the 500 follows, we are being rate limited by @X so we can't immediately follow back but tomorrow is another day 🫡
Love you all
5
4
89
Unfortunately I don't have great news for 6.2 Astra.
I got a containerd image corruption issue on one k3s node, something that should be fairly easy to solve.
Put 6.2 Astra Max Fast to work: 1 hrs of back and forth without solving anything and no clear cause.
Fired up Opus 5.5, resolved in 3 minutes.
What are we doing here?
1
38
Should you need a curated OpenAI models and price catalog you can use our public endpoint: icoretech.github.io/openai-j…
It's the same we use for Codex Pooler.
Open source @ github.com/icoretech/openai-…
1
52
warden-mcp has been a godsend for connecting agents to my keychains, but it can be a bit dangerous with less disciplined models, so perhaps start it read-only. Never had issues with it, though. Open source.
github.com/icoretech/warden-…


