/be-paranoid
a quiet scanner is not a safe app. this command makes the agent plan a security pass from the attacker's side, decide up front what proof would count, and cover each tool's blind spot with another one, before it runs anything.
when to use
- the user types
/be-paranoid, optionally followed by a task ("/be-paranoid review this PR", "/be-paranoid before we launch").
- before merging auth, payments, file upload, admin, or AI-feature changes, even if not asked. say you are doing it.
step 0: pin the task
write one line: what is being checked (repo, branch, diff, feature, or staging URL) and what "done" means to the user. if a live target is involved, check that keelsec.scope.toml lists it and the user has confirmed they own it. if not, the live skills are off for this run and you say so.
step 1: rewrite the task through three lenses
write a short answer to each before running anything. keep each to 3 to 6 bullets.
lens 1, attacker: what would someone hit first?
- list the entry points: routes, forms, uploads, webhooks, CLI args, files the app reads, prompts the AI reads, CI that runs on PRs, MCP servers and skills the agent loads.
- rank the top 3 by "easy to reach" times "bad if it breaks". unauthenticated beats authenticated, money and data beat cosmetics.
- name the one thing you would try first if you wanted in.
lens 2, proof: what shows it is actually safe or fixed, not just quiet?
- for each top entry point, write the evidence that would convince a skeptic: a failing-then-passing test, a lastproof result of
not_reproduced after a fix, a probe set passing 3 of 3, a dependency moved past the fixed version.
- "no findings" from one tool is not proof. write down what that tool cannot see.
lens 3, blind spot: which other field or tool catches what this one misses?
- for each skill you plan to rely on, name its blind spot and the skill that covers it. defaults:
- sinkmark sees code, not config or running behavior. cover with hullscan and tidegate.
- grainseek sees committed text, not whether a key is live. cover with keyprove (with the user's yes).
- depfathom sees manifests, not what ended up in the image. cover with holdsweep.
- static skills cannot tell if a bug is reachable. cover with lastproof on staging.
- seamread is an AI reviewer and can be steered. cover with the deterministic skills plus jibprobe reviewer mode.
- every agent skill trusts its own instructions. cover with bilgecheck.
step 2: pick the skills, order, and pass checks
build a run table. only include skills the lenses called for. local first, live last.
| order |
skill |
why (which lens) |
pass check |
| 1 |
bilgecheck |
blind spot: the agent's own setup |
no high findings in agent files |
| 2 |
grainseek |
attacker: leaked keys are the cheapest way in |
no non-example secrets, examples listed |
| 3 |
depfathom |
attacker: known CVEs are public |
no high or critical without a plan |
| 4 |
hullscan |
blind spot: config and images |
no root final stage, no secrets in ENV |
| 5 |
sinkmark |
attacker: injection at top entry points |
every high sink fixed or justified |
| 6 |
seamread |
proof: the diff introduces nothing new |
zero new high findings, canary not triggered |
| 7 |
jibprobe / evalyard / longwatch |
attacker: the AI feature reads outside text |
all high probes pass 3 of 3 |
| 8 |
portchart, tidegate, probeplate |
proof: what staging really exposes |
scoped only; no unexpected open ports or unauth routes |
| 9 |
lastproof |
proof: is each high static finding real |
each high is confirmed, not_reproduced, or inconclusive with a reason |
| 10 |
ropewalk, squallfuzz |
blind spot: logic bugs and parser crashes |
only if the user asks; time boxed |
| 11 |
keyprove, holdsweep |
blind spot: liveness and the built image |
per-provider yes for live checks |
drillcard is the long form of this table; use it for a full audit rather than a single change.
show the table to the user. live rows (8 to 10) run only with the scope gate passed. if the user is not around, run rows 1 to 7 and list the rest as "not run, needs scope".
step 3: run and check
- run each row, save its JSON, and mark the pass check met or not met right after. do not batch the checking to the end.
- if a row fails its pass check, keep going with the rest, then come back to fixes.
- after any fix, rerun the row that found it, and the row that covers its blind spot.
step 4: merged verdict
merge all findings by dedup_key (kind plus path and line or target, or package plus advisory group, plus CWE). same key means one finding with every source listed and the highest severity kept.
write the verdict in this shape:
verdict: block | ship with fixes listed | ship
checked: <what, which branch or target, when>
top risks (max 5): <finding id, one line, status, proof>
fixed this run: <id, proof of fix>
accepted: <id, reason, who accepted, expiry>
not covered: <languages, folders, live checks not run, and why>
lens notes: attacker first-hit was <x>, its status is <y>
rules for the verdict:
ship needs every pass check met and every high finding fixed with proof or accepted by the user.
- any confirmed high or critical that is open means
block.
- "not covered" is never empty. if you think it is, look again.
also write the merged findings as JSON and SARIF next to the verdict.
stop and scope
- this command plans and runs local skills freely. it never runs a live, fuzz, or pentest skill without the scope gate, never touches cloud metadata addresses, and never uses uncensored or abliterated models.
- it never edits the scope file, never pushes, and never posts results anywhere. it reports to the user.
1---2name: be-paranoid3description: "The /be-paranoid command. Use before any security-sensitive task, review, or release to rewrite the task through three lenses (attacker, proof, blind spot), pick which keelsec skills to run and in what order with a pass check for each, and finish with one merged, deduplicated verdict."4metadata:5 source: https://skillmd.com/skills/codejunkie99/be-paranoid6---78# /be-paranoid910a quiet scanner is not a safe app. this command makes the agent plan a security pass from the attacker's side, decide up front what proof would count, and cover each tool's blind spot with another one, before it runs anything.1112## when to use13- the user types `/be-paranoid`, optionally followed by a task ("/be-paranoid review this PR", "/be-paranoid before we launch").14- before merging auth, payments, file upload, admin, or AI-feature changes, even if not asked. say you are doing it.1516## step 0: pin the task17write one line: what is being checked (repo, branch, diff, feature, or staging URL) and what "done" means to the user. if a live target is involved, check that `keelsec.scope.toml` lists it and the user has confirmed they own it. if not, the live skills are off for this run and you say so.1819## step 1: rewrite the task through three lenses20write a short answer to each before running anything. keep each to 3 to 6 bullets.2122**lens 1, attacker: what would someone hit first?**23- list the entry points: routes, forms, uploads, webhooks, CLI args, files the app reads, prompts the AI reads, CI that runs on PRs, MCP servers and skills the agent loads.24- rank the top 3 by "easy to reach" times "bad if it breaks". unauthenticated beats authenticated, money and data beat cosmetics.25- name the one thing you would try first if you wanted in.2627**lens 2, proof: what shows it is actually safe or fixed, not just quiet?**28- for each top entry point, write the evidence that would convince a skeptic: a failing-then-passing test, a lastproof result of `not_reproduced` after a fix, a probe set passing 3 of 3, a dependency moved past the fixed version.29- "no findings" from one tool is not proof. write down what that tool cannot see.3031**lens 3, blind spot: which other field or tool catches what this one misses?**32- for each skill you plan to rely on, name its blind spot and the skill that covers it. defaults:33 - sinkmark sees code, not config or running behavior. cover with hullscan and tidegate.34 - grainseek sees committed text, not whether a key is live. cover with keyprove (with the user's yes).35 - depfathom sees manifests, not what ended up in the image. cover with holdsweep.36 - static skills cannot tell if a bug is reachable. cover with lastproof on staging.37 - seamread is an AI reviewer and can be steered. cover with the deterministic skills plus jibprobe reviewer mode.38 - every agent skill trusts its own instructions. cover with bilgecheck.3940## step 2: pick the skills, order, and pass checks41build a run table. only include skills the lenses called for. local first, live last.4243| order | skill | why (which lens) | pass check |44|---|---|---|---|45| 1 | bilgecheck | blind spot: the agent's own setup | no high findings in agent files |46| 2 | grainseek | attacker: leaked keys are the cheapest way in | no non-example secrets, examples listed |47| 3 | depfathom | attacker: known CVEs are public | no high or critical without a plan |48| 4 | hullscan | blind spot: config and images | no root final stage, no secrets in ENV |49| 5 | sinkmark | attacker: injection at top entry points | every high sink fixed or justified |50| 6 | seamread | proof: the diff introduces nothing new | zero new high findings, canary not triggered |51| 7 | jibprobe / evalyard / longwatch | attacker: the AI feature reads outside text | all high probes pass 3 of 3 |52| 8 | portchart, tidegate, probeplate | proof: what staging really exposes | scoped only; no unexpected open ports or unauth routes |53| 9 | lastproof | proof: is each high static finding real | each high is confirmed, not_reproduced, or inconclusive with a reason |54| 10 | ropewalk, squallfuzz | blind spot: logic bugs and parser crashes | only if the user asks; time boxed |55| 11 | keyprove, holdsweep | blind spot: liveness and the built image | per-provider yes for live checks |5657drillcard is the long form of this table; use it for a full audit rather than a single change.5859show the table to the user. live rows (8 to 10) run only with the scope gate passed. if the user is not around, run rows 1 to 7 and list the rest as "not run, needs scope".6061## step 3: run and check62- run each row, save its JSON, and mark the pass check met or not met right after. do not batch the checking to the end.63- if a row fails its pass check, keep going with the rest, then come back to fixes.64- after any fix, rerun the row that found it, and the row that covers its blind spot.6566## step 4: merged verdict67merge all findings by `dedup_key` (kind plus path and line or target, or package plus advisory group, plus CWE). same key means one finding with every source listed and the highest severity kept.6869write the verdict in this shape:7071```72verdict: block | ship with fixes listed | ship73checked: <what, which branch or target, when>74top risks (max 5): <finding id, one line, status, proof>75fixed this run: <id, proof of fix>76accepted: <id, reason, who accepted, expiry>77not covered: <languages, folders, live checks not run, and why>78lens notes: attacker first-hit was <x>, its status is <y>79```8081rules for the verdict:82- `ship` needs every pass check met and every high finding fixed with proof or accepted by the user.83- any confirmed high or critical that is open means `block`.84- "not covered" is never empty. if you think it is, look again.8586also write the merged findings as JSON and SARIF next to the verdict.8788## stop and scope89- this command plans and runs local skills freely. it never runs a live, fuzz, or pentest skill without the scope gate, never touches cloud metadata addresses, and never uses uncensored or abliterated models.90- it never edits the scope file, never pushes, and never posts results anywhere. it reports to the user.