Research Detection engineering
Authentication telemetry needs context.
Repeated failures are useful signals for analyst review. They only become meaningful when source, identity, method, time, and subsequent behavior remain connected.
Authentication logs are full of repeatable events: failed passwords, invalid user names, supported methods, successful logins, and source addresses. A parser can count them quickly. The more careful task is deciding what the count does—and does not—mean.
A signal is not a verdict.
A burst of failures from one address can justify a review. It can also result from a user with an outdated client, a misconfigured service, shared network infrastructure, a scanner in an authorized environment, or an event that cannot be resolved from the available source alone.
Defensive tooling should keep that distinction visible. It can report the observation and the reason it crossed a configured threshold; it should not silently turn a pattern into a claim of malicious intent.
Context changes the review.
The following redacted, illustrative fragment contains a source, a sequence of attempted names, an authentication method, and a later event worth reviewing. It is not production telemetry.
2026-08-13T08:14:02Z failed user=admin source=198.51.100.42 method=password
2026-08-13T08:14:07Z failed user=deploy source=198.51.100.42 method=password
2026-08-13T08:14:15Z failed user=backup source=198.51.100.42 method=password
2026-08-13T08:21:43Z accepted user=known-user source=198.51.100.42 method=publickey
The sample does not establish a conclusion. It gives a reviewer concrete questions: are the accounts expected, did the source appear elsewhere, does the successful login belong to a known identity, and what surrounding telemetry exists?
- Source: address, network context, and whether address sharing is plausible.
- Identity: user names, invalid accounts, and later successful authentication.
- Method: password, public key, keyboard-interactive, or another supported event type.
- Time: event ordering and the window in which a sequence occurred.
- Coverage: what the parser supports and what it cannot see.
Thresholds should stay explainable.
Thresholds make large datasets reviewable, but they should travel with the result. If a rule reports repeated failures after a given number of events, the output should show that number. If a password-spray-style heuristic looks for multiple names from one source, the output should expose the names and count that triggered it.
Output for a human reviewer.
Structured output retains the path from raw event to reported observation. A reviewer should be able to see the source, count, affected names, event types, and explanatory rationale before deciding whether the signal needs wider investigation.
Auth Log Analyzer is a public Python CLI that produces JSON and Markdown observations from supported OpenSSH authentication events. Its purpose is to reduce the distance between raw logs and a reviewable finding—not to replace a SIEM, EDR, or human validation.
Make the limitation part of the result.
Context-free thresholds can produce false positives. A log source can omit time-zone information. A mapping can assist a reviewer without proving adversary intent. These are not flaws to conceal; they are conditions the next person needs in order to interpret the output responsibly.