Every technology that makes a website legible to AI agents: what each one is, how widely it is really adopted, what Lumar checks when a site publishes it, and a correct and a broken example of each.
Not adopting an emerging technology is not a failure. Every check is gated on detecting the technology first, so a site that has never published one of these is not assessed on it and appears in no report. What this tool surfaces is a technology that is present and wrong.
Security scope and precision
Security checks are provisional: precision has not yet been measured for this ruleset. A clean row means no listed defect was detected in the evidence read; it is not a security assessment, penetration test or certification. Absence of a technology is never a fault. Partial coverage is reported separately.
What the scanner does
The scanner reads public pages, response headers, advertised metadata and public key sets with time, byte and destination limits. On each target host it guesses conventional discovery paths, including several MCP card paths and OpenAPI paths, and sends unauthenticated POST probes to /ask and /mcp even when no protocol was advertised. MCP probes may include server/discover, a legacy initialize fallback, tools/list and an unknown-method request. A2A endpoint probes send an empty message:send request. Advertised metadata GETs can reach third-party hosts. Signed A2A cards can explicitly name JWS key sets: at most three are read per card, using the shared metadata cache. Known test-key checks match public fingerprints; they do not verify signatures or establish trust in the signer. A separately discovered MCP server card or A2A card can also name a third-party endpoint for the existing protocol POST probes; these requests are unauthenticated. UCP service declarations authorize no extra service POSTs here: we read OAuth provider metadata and advertised A2A cards, and leave service execution unassessed. This is bounded public-evidence collection, not GET-only scanning. Addresses or redirect destinations containing URL user information or nonempty credential query parameters are not requested; the affected check stays unassessed. Caller bypass headers are restricted to the target origin and are removed on cross-origin redirects. The scanner does not authenticate, execute a purchase, replay payments or test runtime defenses.
Severity and authority
Severity is assigned per catalogued check from the observed protection and potential consequence. High includes credential and signing-key exposure and identity-binding failures. Medium includes transport risks, weak published keys and other bounded configuration risks. Low covers limited guidance departures. These examples do not define exclusive categories. Severity is remediation priority, not exploitability or incident confirmation. Each defect lists its authority and CWE; Lumar recommendations are identified separately and do not claim specification conformance. Framework crosswalks are related risks, not proof of complete framework coverage.
Exposure handling and disclosure
The public scanner, API and MCP tool may report the existence of private material or a known published test key on a public domain. Findings name the fault and private member names, never their values. URL user information and credential query parameters are redacted in metric outputs, request trails, fix prompts and newly persisted documents. This redaction is technical enforcement for arbitrary-domain output; domain ownership is not verified. Newly collected study artifacts redact private JWK values and explicit credential fields before persistence; malformed sensitive documents are withheld. Historical artifacts require a separate restricted-access review and redaction before reuse or publication. Do not use a discovered key, publish its value or include it in a ticket. Remove it from the public document and rotate it promptly. Contact Lumar through its contact page to report an exposure or dispute a finding; include the domain, ruleset, flag, scan time and redacted evidence. An automated finding does not contact the owner or establish a disclosure deadline. Lumar-platform access controls and retention are outside this scanner's assessment.
Content
Once an agent has a page, can it read it cheaply and correctly? Read the category.