What this actually looks like on a Tuesday afternoon
Ten scenarios, one per crew member, each walked from the triggering moment to what was recorded and what it came to. Two are things the product does today. The rest are the workflows the roadmap modules are specified against — labelled, so you always know which is which.
M1 GenAI Prompt DLP
A finance analyst pastes 200 customer rows into ChatGPT
Available for POC
The single most common way sensitive data reaches a public model — and the one a network control cannot see, because the session is TLS to a legitimate business domain.
Trigger
Paste into the ChatGPT composer. The browser calls the local connector and blocks before the content reaches the page.
Decision
Verdict service: 47 NRIC checksum matches, plus 12 Exact Data Match hits against the enrolled customer-account table. Crown-jewel EDM is a high-precision class, so the rule action is hard-block.
Human action
Policy resolves. The user is in SG-FIN-Analysts, which is at enforce stage, so no downgrade applies. No exception is active. Final action: block.
Evidence
One immutable event per capture, verdict, warn, override, block and dispute
Limitation
It cannot escalate its own verdict. A model-based classification does not hard-block by default, because at real-world classification precision a block-by-default posture produces false blocks at fleet scale. That posture was considered and rejected.
Current deterministic capability; the browser demonstration verifies behaviour, not production readiness.
The attachment that was one hidden worksheet from a notifiable breach
Design-partner roadmap
Not an attacker. A relationship manager in a hurry, on a Friday afternoon.
Trigger
Q3_review.xlsx is attached and sent to a client contact. The message is held pre-delivery; the sender's client shows it as sending.
Decision
The workbook is extracted in the sandbox. A third worksheet, hidden and not visible in the sender's view, contains 1,340 rows belonging to a different client.
Human action
62 Exact Data Match hits against the enrolled customer table. The recipient is known-external and legitimate — but not entitled to that data.
Evidence
M0-schema events for every analysis, hold, release and block
Limitation
It cannot confirm a compromise or authorise a release. It flags, explains, and hands the decision to a person.
Roadmap workflow specification; not built and not presented as a product simulation.
Two findings in one diff, produced by two different kinds of authority.
Trigger
A pull request opens. The Git host adapter delivers the diff, the manifests, and the resolved dependency tree.
Decision
Software composition analysis flags a transitive dependency introduced by the bump, matched against the local OSV pack, carrying a known-exploited entry. This is authoritative: a scanner said it, and it is reproducible.
Human action
Reachability analysis places the vulnerable function on a path called from the file-upload handler, which is internet-facing. That combination moves it above 40 other open findings.
Evidence
Findings with file and line evidence, provenance, and reviewer disposition
Limitation
It is not the source of truth for a vulnerability, and its prioritisation is a recommendation rather than a risk decision.
Roadmap workflow specification; not built and not presented as a product simulation.
One analyst on shift, an EDR alert on a finance workstation, and six systems holding one piece of the answer each.
Trigger
The analyst asks the question in plain language rather than opening six consoles.
Decision
A cited timeline returns: the EDR alert; a VPN session on the same host from an unusual ASN 40 minutes earlier; two failed Kerberos pre-authentications followed by one success; a DNS lookup to a domain first observed six days ago. Each step links to its source event.
Human action
Proxy logs for an 18-minute window are missing — that connector was degraded. The copilot names the gap instead of reasoning across it, because a confident narrative over a hole is worse than no narrative.
Evidence
An invocation record per AI call: model and adapter version, prompt-template version, retrieval-source identifiers, requesting identity, tool calls, policy context, output disposition, latency, reviewer outcome
Limitation
It cannot establish that an intrusion occurred. Model output alone is not proof of compromise.
Roadmap workflow specification; not built and not presented as a product simulation.
The finding is about a folder. It is deliberately not about a person.
Trigger
An approved scan of one departmental file share, agreed in advance with the data owner.
Decision
11,400 files containing regulated personal data. 3,900 of them older than the retention schedule permits — an over-retention exposure that exists whether or not anyone ever misuses it.
Human action
One folder inherits permissions from a security group with 340 members, of whom 12 are actually in the department. Deterministic graph analysis, not an inference.
Evidence
Classification suggestions with the located material and its owner
Limitation
It cannot conclude intent. An inference about a person is a hypothesis with evidence attached, never a determination.
Roadmap workflow specification; not built and not presented as a product simulation.
Thirty-four controls, two days instead of six weeks
Design-partner roadmap
The value is not the writing. It is knowing which twenty-nine you can answer and which five you cannot.
Trigger
The regulator asks for evidence against 34 controls, with a response window that assumes the organisation already knows where everything is.
Decision
29 controls map cleanly to artifacts by identifier, each with a citation to the source document and version.
Human action
3 map to artifacts older than the control's own review cycle. They render as stale rather than being submitted as current — which is the difference between a weak answer and a false one.
Evidence
Cited audit responses traceable to source document and version
Limitation
It cannot assert that you are compliant. It creates reviewable gaps; it does not certify.
Roadmap workflow specification; not built and not presented as a product simulation.
The question that decides whether an observability tool is an instrument or a decoration.
Trigger
The projection shows unsanctioned AI destinations up 40% week on week. The figure renders with its watermark and projection version attached.
Decision
By OU, the increase concentrates almost entirely in one department. Low-count suppression holds; no individual resolves.
Human action
Three of the top five destinations sit in the suspected-AI bucket — real traffic, LLM-shaped, matching no catalog entry. The dashboard says suspected, on the tile, not in a footnote.
Evidence
Shadow-AI inventory by destination, AD group and OU, trended over time
Limitation
It cannot see what proxy and DNS cannot see. Local and on-device models of the Ollama class are invisible to this telemetry; that gap is recorded, not papered over.
Roadmap workflow specification; not built and not presented as a product simulation.
The question an auditor asks that most products cannot answer, because an audit log without integrity protection is worth nothing to a regulator.
Trigger
The auditor exports events #000001–#004417 with the signed chain head, using their own read-only role. No administrator is involved in producing the export.
Decision
They recompute the chain independently. The platform is not asked to verify itself — that would be the whole problem.
Human action
Recomputation matches until #000414, then diverges. One event's payload had been altered directly in the database, bypassing the application.
Evidence
Versioned event schema with actor as SID + sAMAccountName, machine as AD computer object, policy version in force, and evidence reference or none
Limitation
Not applicable — there is no inference in the platform core.
Current deterministic capability; the browser demonstration verifies behaviour, not production readiness.
Seven of these describe modules we have specified and not built. We publish them because
the specification is the honest thing to show — actor, trigger, happy path, exceptions,
terminal state, and the audit evidence it must leave. A simulated demonstration of unbuilt
software would tell you less and imply more.
Sentry's decision path and the audit chain are not descriptions. The detection logic and
the hashing both execute in your browser, with no network request, so you can check the
behaviour rather than take our word for the case.
