Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Secure Exa API keys, implement content moderation, and manage domain restrictions. Use when securing API keys, auditing Exa security configuration, or implementing content safety filtering. Trigger with phrases like "exa security", "exa secrets", "secure exa", "exa API key security", "exa content moderation".
.claude/skills/jeremylongshore-exa-security-basics/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 20% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 18% | 0% |
| case-18 | ✗→✓ | ▲ Improved | -36% | 0% |
| case-22 | ✗→✓ | ▲ Improved | 0% | 0% |
| case-04 | ✓→✗ | ▼ Worse | 46% | 0% |
Threat-model Exa credentials, query intent, retrieved web content, generated output, and retained operational evidence as separate trust boundaries. Treat credentials, queries, retrieved content, generated output, spend, and destructive state as separately governed boundaries.
Exa is SOC 2 Type II certified and offers enterprise controls such as Zero Data Retention and HIPAA enablement. HIPAA mode is request-scoped, supports only eligible Search and Contents cache-only retrieval, and rejects summaries or live freshness. Regional access restrictions can yield Cloudflare block pages rather than Exa JSON.
For normal REST work, inject EXA_API_KEY from an approved server-side secret manager and send it only as Authorization: Bearer to the configured first-party Exa API host. Team Management service keys, hosted MCP OAuth or enterprise managed authorization, and payment-protocol calls are separate trust models. Never print, commit, place in a URL, or expose a credential to an untrusted client.
Use Read, Glob, and Grep to inspect repository code, configuration, fixtures, and evidence. Use Write and Edit only for approved implementation or documentation changes. Do not call Exa, run paid research, create or alter a Monitor, Webset, Agent run, Batch, team, member, API key, budget, webhook, or deployment merely because this skill was invoked.
Require an accountable owner before live queries involving sensitive intent, production credentials, spend or rate-limit changes, forced live crawling, generated summaries, external delivery, deployment, member or key changes, schedule creation, or destructive cancellation, stopping, deletion, or revocation. Read-only repository inspection and synthetic offline validation do not authorize live vendor actions.
Return the operation scope, environment, team and product surface, authorization class, contract and policy decisions, deterministic validation results, content-free identifiers, status and cost counts, risks, cleanup or rollback state, and a concise pass or fail receipt. Exclude credentials, raw queries, prompts, presigned URLs, retrieved content, generated output, and customer-derived data unless separately approved.
Rerun the smallest relevant deterministic test, compare actual behavior with the requested outcome and current first-party contract, verify sensitive fields are absent from evidence, and confirm deadlines, terminal state, downstream retention, and rollback before reporting success.
Review the dated first-party evidence map before relying on any endpoint, parameter, search type, price, limit, beta, compliance, identity, retry, or lifecycle claim.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 17,469 | 12,995 | -26% | 1 | 1 | 0% | 2,801 | 3,375 | +20% | 0 | 0 | — |
case-02 | fail→fail | 15,236 | 13,127 | -14% | 1 | 1 | 0% | 2,980 | 3,636 | +22% | 0 | 0 | — |
case-03 | pass→pass | 17,523 | 33,541 | +91% | 1 | 1 | 0% | 3,386 | 3,844 | +14% | 0 | 0 | — |
case-04 | pass→fail | 8,912 | 7,480 | -16% | 1 | 1 | 0% | 1,765 | 2,574 | +46% | 0 | 0 | — |
case-05 | pass→pass | 15,854 | 13,881 | -12% | 1 | 1 | 0% | 2,431 | 4,099 | +69% | 0 | 0 | — |
case-06 | pass→pass | 8,585 | 9,833 | +15% | 1 | 1 | 0% | 1,692 | 3,051 | +80% | 0 | 0 | — |
case-07 | pass→pass | 4,239 | 2,876 | -32% | 1 | 1 | 0% | 847 | 1,854 | +119% | 0 | 0 | — |
case-08 | pass→pass | 9,614 | 4,487 | -53% | 1 | 1 | 0% | 1,458 | 2,145 | +47% | 0 | 0 | — |
case-09 | pass→pass | 6,884 | 3,449 | -50% | 1 | 1 | 0% | 979 | 2,104 | +115% | 0 | 0 | — |
case-10 | pass→pass | 10,277 | 3,851 | -63% | 1 | 1 | 0% | 1,606 | 2,144 | +33% | 0 | 0 | — |
case-11 | pass→pass | 11,403 | 8,280 | -27% | 1 | 1 | 0% | 2,161 | 3,030 | +40% | 0 | 0 | — |
case-12 | fail→pass | 16,373 | 8,966 | -45% | 1 | 1 | 0% | 2,436 | 2,867 | +18% | 0 | 0 | — |
case-13 | pass→pass | 8,435 | 4,948 | -41% | 1 | 1 | 0% | 1,501 | 2,228 | +48% | 0 | 0 | — |
case-14 | pass→pass | 7,063 | 3,541 | -50% | 1 | 1 | 0% | 1,302 | 2,113 | +62% | 0 | 0 | — |
case-15 | pass→pass | 11,476 | 3,854 | -66% | 1 | 1 | 0% | 1,962 | 2,058 | +5% | 0 | 0 | — |
case-16 | pass→pass | 14,654 | 3,790 | -74% | 1 | 1 | 0% | 2,176 | 2,089 | -4% | 0 | 0 | — |
case-17 | pass→pass | 8,199 | 8,767 | +7% | 1 | 1 | 0% | 1,518 | 2,809 | +85% | 0 | 0 | — |
case-18 | fail→pass | 17,695 | 3,477 | -80% | 1 | 1 | 0% | 3,228 | 2,073 | -36% | 0 | 0 | — |
case-19 | pass→pass | 11,247 | 6,409 | -43% | 1 | 1 | 0% | 1,900 | 2,389 | +26% | 0 | 0 | — |
case-20 | pass→pass | 6,888 | 6,698 | -3% | 1 | 1 | 0% | 1,457 | 2,407 | +65% | 0 | 0 | — |
case-21 | pass→pass | 11,605 | 5,940 | -49% | 1 | 1 | 0% | 1,851 | 2,697 | +46% | 0 | 0 | — |
case-22 | fail→pass | 11,601 | 3,258 | -72% | 1 | 1 | 0% | 2,091 | 2,100 | +0% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +14 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
The publisher has shipped newer versions since this run, so these numbers describe v1, not the version currently listed.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.