A webhook signature and permission to act are different checks
Separate webhook authenticity from authority, event context, retries and approval for an AI action.
Article language: English

In this article
The message came from the right service and its signature verified. An AI agent read the comment “you can publish” and published a new catalogue version. The commenter was a customer asking for a preview, not a person authorised to approve release. The delivery sender was genuine. Permission to publish arose only in the automation's interpretation.
This is a hypothetical process at a company collecting catalogue feedback in an issue system. A change in a case triggers a webhook: a message a service sends to a specified receiver after an event. The agent should organise comments and prepare a proposed next step. Publication requires a separate decision by the material's owner.
Accepting such a process requires more than asking whether the webhook works. Separate proof of a delivery's origin and unchanged content from deciding what may be done on that basis. A signed message can accurately report an action by someone unauthorised for your subsequent action. It can also arrive again or concern an outdated case state.
What does a signature actually confirm?
Validate a webhook signature using the particular provider's mechanism. The process owner should know which component checks it, with what configuration and what happens on failure. Assessing cryptographic implementation correctness belongs to an appropriately qualified implementer.
GitHub describes signature validation before further processing a delivery. In its HMAC mechanism, based on a shared secret key, the signature depends on the webhook secret and message body. The aim is to check origin and that the delivery body has not been altered. An inference about origin depends on protecting the secret and correctly associating it with the sender. A valid signature alone does not establish that a delivery is fresh.
This is a useful boundary to draw in a company process. An issue service can accurately report that someone added a comment. That does not make the comment an instruction from the catalogue owner. Trust in the delivery channel and the content author's authority need separate assessment. AI should not infer missing approval from a sentence's polite tone.
When comparing schemes, list which fields the signature check covers. Do not assume every header and identifier is protected identically to the body. The implementer's evidence should identify the scope consistent with the service documentation. A “valid signature” report must not add confirmation of every request element by itself. This matters especially when some data are later used to choose a process or recognise a duplicate.
Prepare a simple responsibility map. First the receiver gets a delivery. Next come appropriate authenticity and required-data checks. Only an accepted event may proceed to the intended interpretation. Before a consequential action, process rules still apply: the right case, role, state and necessary approval. The steps may operate in different services.
Check that an invalid delivery is stopped before its content reaches the model. If an agent has already begun retrieving documents or preparing actions, a later “invalid signature” label does not describe the whole effect. Acceptance needs observation at the relevant point in the process, not just the final server message.
Do not put the signing secret in an AI instruction, acceptance report or ordinary team message. Evidence needs configuration identifiers, result and owner, not a value enabling impersonation of a valid delivery. If a test cannot be shown without revealing the secret, prepare a better documentation method.
A valid delivery may still not fit the task
A source system sends different event types. Adding a comment, editing a description, changing an assignee and approving material need not mean the same thing. If a receiver runs the same flow for every message, a valid signature cannot save the process's meaning. First determine whether the event should be handled at all.
GitHub recommends checking event type and action before processing. This is separate from signature validation; the service itself can add new types and actions.
At the fictional company, a customer comment goes to feedback review. Release approval belongs to another stage. Record how the system recognises each and who may approve. “I accept” in an arbitrary field need not be approval. A model may summarise such a statement, but should not promote its authority itself.
Catalogue case K-17 also changes version. A customer comments on its preview. The agent records a proposed correction and an editor approves a prepared version. Before publication, someone adds another comment and changes the document. The channel still works, but earlier approval does not automatically cover new content. The receiver must preserve the relationship between event, material considered and permitted operation. The team should know whether the process stops for another decision or performs another previously agreed step.
Include a correctly delivered event from another project or account in acceptance. Its content may be entirely reasonable but still not belong to the catalogue covered by the automation. Establish where the proper context comes from and where it is enforced. Writing “handle only our company” in a model instruction does not replace a resource-association check.
AI integration permissions (in Polish) concern later read and write capabilities. A receiver may accept a message even though its execution account should not publish the catalogue. The reverse is also possible: a broad account can technically publish, but the current event does not justify it. Both boundaries belong on one map.

A more elaborate connection needs a separate account of the token recipient and MCP service boundaries (in Polish). A valid incoming-message signature does not explain permissions the intermediary will later use. One positive check at the beginning must not make all later stages approved by default.
The same message may return
A connection can fail or an operator can request redelivery. Accepting a second correctly signed message is not automatically new approval for a second effect. Determine whether it is the same event and connect it with earlier state. Input handling meets duplicate and retry control (in Polish) here.
In the model described by Standard Webhooks version 1.0.0, the signature covers identifier, attempt time and body. The specification separates checking time from recognising a repeat. This is one scheme, not a description of every provider.
For your solution, ask which identifier remains stable on retry and how long the system can connect it with prior execution. Do not copy a protection period from a random documentation example. The process may need later manual redelivery, not only two attempts close together. The expected period should fit actual operations.
Also distinguish event time from delivery-attempt time. Standard Webhooks describes that difference. A new attempt may bring an old event that should not overwrite current status. A late status-changing message (in Polish) needs a different decision from an ordinary duplicate. In our catalogue an old comment must not by itself restore a withdrawn version for publication.
During a test record both outcomes: delivery acceptance and the action state in the destination system. One entry in the receiver's history does not prove one publication if several downstream services retry independently. Equally, two deliveries need not mean two effects. An operator should distinguish those counts without reconstructing the entire implementation from the author's memory.
GitHub describes quickly acknowledging receipt and processing further work in a queue. A transport response can therefore appear before downstream work completes. Do not present it as evidence of catalogue publication.
Our case needs at least recognisable records of receipt, scope decision and operation result. They need not form three separate panels. They must let an operator distinguish a comment waiting for an editor from publication completed in the destination. If a later stage fails, retrying the entire delivery should not be the only response. First reconcile what has already happened and what still needs action.
A signature does not turn a comment into an agent instruction
A catalogue comment may contain a page address, attachment or description of an action. It remains content from the comment author. If an agent should summarise it, the agent should not itself assume it may visit arbitrary addresses, send documents there or apply new instructions to other cases. The channel signature does not expand the model's task.
Links and images in an AI response (in Polish) are another potential data-exit path. An input review should determine what the agent does with an external link found in an event. Reading text and making a network request have different effects. Use harmless markers and controlled resources in the test.
A process owner need not implement signature validation personally to ask good questions. They should be able to identify where message acceptance ends and interpretation and action begin. Name each missing piece of evidence at that point. A general claim that “the webhook is secure” does not decide which boundary was actually checked.
A small test should show different reasons to stop
Prepare an authorised environment and fictional case with the implementer. The normal flow should reach an agreed effect, such as a draft reply to a comment. Later cases should stop for different reasons: invalid origin check, unsupported type, wrong project, repetition of an already reconciled event and missing publication approval.
Five identical error messages are not enough. For each case check whether the downstream effect did not occur and whether the operator has the right next step. A signature error needs different handling from a valid comment waiting for an editor's decision. The model should not bypass a stop through another tool or turn every refusal into “try again later”.
For a configuration change, also prepare a return-to-work test. Confusing the test-environment secret with the production secret can stop valid deliveries. The operator should not respond by disabling validation. They need contact with the configuration owner and a way to find and reconcile backlogged cases. Restoring a green status alone does not say what happened to messages rejected during the interruption.
After removing the cause, test resuming the same case. An editor's approval of one catalogue version does not mean approval of every future change in the issue. Preserve the link between the decision, correct material and scope. The log needs only the necessary trace, without copying full content or secrets into another location.
At Aeneia Audit, the operational use of AI and a map of such checks can be reviewed. Specialist implementation and security review requires a separate scope, appropriate expertise and tests. A company trial should produce a recognisable decision: the delivery is accepted, the comment belongs to this case and publication still waits for a person authorised to approve it.







