Job search
Security Analyst Portfolio: Turn a Public Log Dataset into an Investigation Case
Original synthetic logs actually queried in Python/SQLite support a complete report with competing hypotheses and unconfirmed transfer scope.
A security analyst portfolio becomes useful when a reviewer can follow the evidence from a permitted log source to a timeline, a supported judgment, and the questions still unresolved. A public dataset is a starting point, not proof that you discovered a real attack. Show how you handled the records, which conclusion they support, and which questions remain open.
This guide demonstrates the workflow with original synthetic logs that were actually queried locally on October 21, 2026. It does not claim to analyze a downloaded public dataset or a real organization’s activity. LandOffer publishes this guide and offers job-search tools. No external system was scanned, no private log was uploaded, and no attack, account change, or incident-response action was performed.

Use LandOffer’s recently posted jobs to identify the SOC analysis responsibilities your portfolio should demonstrate. Choose an investigation question you can answer with the available records.
Establish provenance before searching for suspicious patterns
For a public dataset, record the publisher, exact file or revision, stated collection context, usage terms, and any labels supplied by its author. Preserve those distinctions when sharing your analysis. A dataset labeled as a simulation should not become an account of an intrusion you personally investigated.
The Security Datasets project describes malicious and benign datasets for analysis and research and displays an MIT license reference. It is one example of a documented source to inspect. That project-level description does not replace checking the particular file, its provenance, and applicable notices. This article did not download or execute its datasets, notebooks, or simulation material.
For a small teaching case, you can create an original synthetic log and label its origin clearly. It lets you expose data-quality problems without publishing anyone’s identifiers. It demonstrates query and reasoning skills, while leaving realism and operational experience unproven.
NIST SP 800-61 Rev. 3, published in April 2025, places incident response within cybersecurity risk management and supersedes Revision 2. Its publication page supplies useful current context. A small portfolio report is not an implementation of the entire publication or a certification of an incident-response program.
Worked asset one: a finite log investigation you can rerun
The original sample contains ten JSONL rows for a fictional application. The declared fields are event ID, explicit UTC event time, event type, account label, source label, optional session, and optional request ID. Account and source names are opaque teaching identifiers, not people or network addresses.
The three event types are login_failed, login_succeeded, and report_download_requested. Their meanings are limited to what the authored schema states. A download request is not a completed transfer. A successful login event is not proof that the person using the session was authorized.
The input deliberately includes one exact duplicate, one invalid timestamp, one missing account, and records stored out of chronological order. The processing script reads the source, records its hash, validates the timestamps, separates the duplicate and quarantine records, and inserts accepted events into an in-memory SQLite database.
The actual result was ten raw rows, eight valid unique events, one duplicate row, and one quarantined timestamp record. The source file remained unchanged. The two saved fresh-process executions produced matching analysis-output hashes. The observed environment was CPython 3.12.14 with SQLite 3.53.4; the project used only existing standard-library modules and installed nothing.
From the supplied demo directory, the repeat command is:
python3.12 investigate.py > replay.json
Use your selected available Python 3.12 executable and record its version. The actual local executable, platform, commands, source hash, script hash, and output hashes are saved in the run receipt. Other environments were not independently tested.
Python’s sqlite3 documentation documents in-memory connections and binding query values with placeholders. This project uses those mechanics for local analysis; it does not connect to a production database or a commercial SIEM. Its SQL and parameter values are included in the analysis output.
Build the timeline without turning missing fields into facts
The focus query is deliberately narrow:
SELECT * FROM events
WHERE account=? AND source=?
ORDER BY event_time,event_id
The bound values are acct-17 and source-A. The completed result is:
- 14:01:00 UTC, L01: login failed for acct-17 from source-A.
- 14:01:20 UTC, L02: another failed login with the same recorded account and source.
- 14:01:45 UTC, L03: a third account-attributed failed login.
- 14:02:00 UTC, L04: login succeeded for that account and source, with session S17.
- 14:02:30 UTC, L05: a report download was requested in S17, request R17.
All times are February 3, 2026. Event IDs and original file positions remain available so a reviewer can trace the result. L05 appears on source line three, before L02 on line four; sorting file lines would therefore give the wrong event sequence. File position is not a real ingestion timestamp, and this constructed order does not demonstrate an actual delayed-log problem.
The script preserves L08, a failed login at 14:01:30 from source-A with no account value, in a separate query. It does not assign that event to acct-17 merely because the source matches. Doing so would create a fourth account-attributed failure unsupported by the record.
The exact duplicate is another copy of L01. Counting it as a separate occurrence would also inflate the sequence. L09, with an invalid timestamp, stays in quarantine with its source position and reason; it is not silently discarded or placed at an invented time.
Two other valid events remain context: L06 records a success for acct-18 from source-B at 14:00, and L07 records a failure for acct-99 from source-C at 14:05. The exercise does not label either event normal or malicious without evidence.
Before interpreting that sequence, inspect what its ordering means. The accepted fixture times are explicit UTC values normalized to the same fixed-second representation. The query orders those event times and then event IDs, making the output deterministic for this sample. If two records shared a timestamp, the event-ID tie-break would provide a repeatable display order, not proof that one action caused the other. Another dataset could require different precision and clock handling.
The source labels are equally limited. Source-A is a recorded category in the invented application schema; it is not an IP address that was enriched, a device that was verified, or a person who was identified. The session label links fields under the teaching assumptions, but no authentication service was contacted to validate it. Preserve those assumptions in the report so a reader can distinguish successful data processing from verified real-world attribution.
Compare hypotheses rather than declaring a compromise
The supported observation is a sequence of three account-attributed failures followed by a success and a later session request. That pattern can motivate further investigation. It does not establish credential theft, a particular attacker, or unauthorized access by itself.
One hypothesis is unauthorized credential guessing followed by session use. Another is a legitimate user retrying credentials and then requesting a report. Both are consistent with these sparse records. There is no user confirmation, authentication-factor context, device context, baseline behavior, or access-policy evidence to distinguish them.
The request also lacks a server response, completion event, byte count, and content information. Calling it “data exfiltration” would skip several unsupported steps. Correlating the request to S17 identifies a relationship in the log fields, not a verified human identity or a completed transfer.
A useful analyst report can therefore be inconclusive while still demonstrating careful work. The reviewer can inspect the query, see the gaps, compare the explanations, and understand which evidence you would request next. Use a case title that describes the observed sequence without declaring a compromise.
Worked asset two: a completed investigation case report
The project includes this original report, labeled INV-03 — Account-attributed authentication sequence and a later request. It describes the actual local analysis of invented records, not a real security incident.
Status and source: Inconclusive teaching investigation. Ten original synthetic JSONL rows covering February 3, 2026, 14:00–14:05 UTC. Source, script, and analysis hashes identify the files used. No public dataset is claimed to have been analyzed in this example.
Question: What does the authentication sequence support, and does the later request establish unauthorized data access?
Processing: Eight accepted unique events, one exact duplicate, and one invalid-timestamp quarantine. The account focus contains L01–L05. L08 remains unassigned; L06 and L07 remain context. Two actual local executions produced identical output without changing the source.
Evidence-supported finding: Three recorded failures for acct-17/source-A precede a success and one download request associated with S17. The statement is about recorded events. It does not establish who acted, whether the access was permitted, or whether a transfer completed.
Hypotheses: Unauthorized credential guessing and legitimate credential retries are both possible interpretations. The current evidence does not select one. No actor attribution or confirmed compromise is assigned.
Unconfirmed scope: Authentication-factor behavior, device identity, account-owner context, access policy, request response, transferred bytes, report contents, and complete event coverage. Correcting the quarantined timestamp could affect the analysis, so the gap remains visible.
Next evidence to request in a real authorized investigation: Relevant authentication and session context; appropriate account-owner confirmation through an approved process; server response or completion records; applicable access-policy information; source clock and timezone validation; and an upstream correction or parser review for L09. These are proposed actions, not evidence collected during this exercise.
Ownership and disposition: The sample analyst owns normalization, queries, reasoning, and the report. Real authentication or application owners would need to supply additional authorized context. No one was contacted, no account was blocked, and no credential was reset. Retain the case as bounded portfolio evidence, not a production incident or a claim of data theft.

Package the case so a reviewer can inspect the reasoning
Keep the source description, field definitions, script, query text, output, and report together. Put the runnable command and tested environment before the conclusion in the project README. Include the exclusions and quarantine reasons; a reader should not need to reverse-engineer why some events disappeared from a chart.
For this example, the README states the synthetic origin, tested environment, no-install dependency boundary, expected counts, and interpretation limits. The analysis output includes SQL, bound values, all accepted events, the account focus, unassigned records, and session requests. Assertions check the counts, focus IDs, unassigned separation, timeline endpoints, original file position, and unchanged source.
Those checks make the finite exercise reproducible. They do not certify an ingest parser for arbitrary malformed logs or prove that its assumptions fit another dataset. For a public file, adapt and document the schema deliberately rather than treating this sample’s event names as a universal standard.
Hashes identify the reviewed artifacts; they do not make the underlying log authentic. Matching repeated outputs shows what happened in these two runs, not that every future environment or every investigation will behave identically. Keep those distinctions in the portfolio summary.
Explain your work without claiming an attack you did not investigate
A truthful portfolio description is: “Built a local Python/SQLite investigation over original synthetic authentication records; preserved duplicate and invalid-time handling, correlated an account/session sequence, and documented two competing hypotheses with unresolved transfer scope.”
An unsupported description would be: “Detected an attacker stealing reports and contained the breach.” This exercise contains neither confirmed theft nor containment. It also does not demonstrate production SOC access, real-user verification, threat attribution, or use of a commercial platform.
If you later analyze an authorized public dataset, state the exact source and what its labels mean, preserve its usage notices, and separate your observations from the publisher’s scenario. Never upload private workplace logs merely to make a portfolio look more realistic. Use the evidence you are permitted to share and make its limitations visible.
Find a relevant role through LandOffer’s current listings, then include the repeat command, traceable events, and explicit unknowns alongside your investigation conclusion.


