nano v2.0.32
nano v2.0.32: Hunts for Self-run, pivt page awareness with actions you confirm, dataset=signals in Search, queries shown as nPL, no lost events across upgrades and backups, and an announced upgrade path for open 0.1.702 OCSF installs.
v2.0.32 includes everything in v2.0.31. It includes security fixes: pivt no longer changes ingestion settings on its own, and risky AI case actions need confirmation by default.
To upgrade, follow Upgrade a self-hosted deployment and the Upgrade notes below. v2.0.20 is a required stop: installs on a release older than v2.0.20 upgrade to v2.0.20 first.
New
Hunts (Self-run)
Hunts are scheduled threat hunts that AI agents run against your data on the nano server. They are in the new Hunts page at /hunts. See Hunts.
- Server runner. Hunts run on the server with the AI provider you configure in nano. They don't need the Agent Lane switch, listener or allow-list that external agents use. nano Desktop can still run hunts too.
- A pipeline per run. Each run first checks which sources and fields you have, then runs the hunt's pivots as fixed queries with no model, and only then hands the results to the investigator. A pivot whose sources are absent is skipped, and a hunt with nothing to run ends without a model call.
- Findings, not guesses. The investigator files each candidate as a lead, benign or a duplicate, including duplicates of a detection that already fired. It remembers baselines, notes and suppressions, each tied to evidence it read.
- An evidence-only reviewer. A second agent checks each lead. It must record its verdict before it can read the investigator's reasoning, and it must cite events it read, including the finding's own. The reviewer is the only agent that can escalate.
- Escalation. Each hunt escalates in one of three modes: off, notify (through your existing webhooks: generic, Slack, Teams, PagerDuty and email) or case. Case mode opens a case marked as needing analyst acknowledgement. It unlocks only when the reviewer has a proven track record. See Escalation.
- Promote to case. Promoting a finding opens the new case with the finding's evidence events, entities, a summary, the queries as nPL, a timeline and next steps.
- Agent detections. A hunt can write detection rules that the detection engine runs on schedule with no AI. They run in Live mode only and never alert. A rule must pass a 7-day backtest first, and each hunt has a cap on how many it keeps. When one fires, the reviewer triages it. They retire on their own after 30 days, or after 14 days with no trips.
- Campaigns and actors. Analysts can track known attackers as campaigns with trip conditions. Hunts recheck active campaigns cheaply on each run. The investigator can propose a campaign, but only an analyst can turn it on.
- Playbooks and templates. Hunts are built from playbooks with a hypothesis and pivots. Pivots are written in nPL over the events, signals or risk datasets. Built-in templates include Detection signals, DNS tunneling and an intrusion chain. Hunt files in a repository can now set
risk_weightanddefault_severity. - Evals. The reviewer model is replayed against findings your analysts resolved, on demand or weekly. A new reviewer model needs its own passing eval before case mode applies to it.
- Budgets and retention. Every run has token, turn, tool-call and time budgets, and holds back a final turn so findings are filed even when the budget runs out. Old runs, dismissed findings and lapsed suppressions are cleaned up on a schedule.
pivt knows the page you are on
pivt now reads what you are looking at on Search, Alerts, Alert detail, Matches, the Rules list, the rule editor and Dashboards, and can propose actions there:
- Ask "what am I looking at?" and pivt answers from the page's own state.
- Anything that changes state shows a card with a preview. Nothing happens until you click Apply. Read-only actions, like filtering or a search pivot, are one click.
- Apply runs the action with your own session and permissions, and the audit log records it as you, marked as coming from pivt. pivt only offers actions you have the permissions for.
- Examples: acknowledge or close selected alerts (where your edition triages alerts on that page), promote, demote, pause or resume a rule, put a proposed query in the rule editor without saving it, and add or edit an nPL panel on a dashboard. pivt never overwrites a newer version or your unsaved edits.
Search
dataset=signals. Search the detections that fired: real-time detections and scheduled rule findings together, with the rule, severity, risk score, risk entity and the matched event. Choose Signals in the dataset selector, or start the query withdataset=signals. You need thedetections:viewpermission. Detection rules can't read this dataset.- Datasets on dashboards. An nPL dashboard panel can pick its dataset (Logs, Traces, Metrics, Risk or Signals) in the panel editor.
- Queries shown as nPL. Where nano shows a query that an agent ran as SQL (hunt runs, playbook pivots, campaign conditions, notebook AI queries and the Agent Lane query audit), it now shows the nPL version first, with Open in Search. The SQL is behind a toggle. A conversion that isn't exact is labeled, and SQL that can't be converted says why. On Self-run, Try AI conversion asks AI for one, and its result is marked as approximate.
- Subsearches run their whole pipeline. In
x IN [search … | stats … | where … | head N | return f], every stage now runs. Before, only the search andreturnran, and the stages in between were ignored. See the upgrade notes.
Upgrade tools
- An announced upgrade path for open 0.1.702 OCSF installs. Re-running the installer on an open 0.1.x install now says it is an upgrade to 2.x, explains what changes and asks you to confirm. In a terminal it also offers a backup (a PostgreSQL dump and a copy of
.env). The backfill tools ship with the release:scripts/nano ocsf-backfillandscripts/nano ocsf-legacy-drop. - Readiness report.
scripts/nano upgrade-readiness(orscripts/nano-upgrade-readiness.py, which also runs on 0.1.x) writes a read-only report of your install: version, layout, data volume per day, the disk the OCSF backfill needs, content counts, and how each ingest path is used.exportwrites your content (parsers, rules, routing, saved searches, dashboards and lookup definitions) without event data or secrets.export --encryptencrypts it for nano support. - Open to Self-run in place.
scripts/nano convert --to enterpriseconverts an open 2.x single server to nano Self-run at the same release. It checks first (--dry-run), takes a backup, verifies the result, and--rollbackrestores the open install. scripts/nano cluster(Self-run). From a workstation, create a role-split cluster on GCP, AWS or Azure sized for your daily volume, then upgrade, scale or delete it. It can set up managed HTTPS and write Terraform instead of creating resources.
Improvements
- No acknowledged event is lost across upgrades and backups (single server). Vector now waits for the API to render the new release's pipeline before it accepts events, and a change to a sink's disk buffer restarts Vector in a controlled way instead of a hot reload. Upgrades,
scripts/nano backup, restore and convert close ingest, wait until Vector's buffers are empty and every event has reached ClickHouse, stop Vector, then reopen ingest. While ingest is closed, nginx answers ingest requests with503andRetry-After, and Vector's own ports refuse connections. Senders that retry or buffer send again after it reopens. - Rejected ClickHouse batches are kept (Vector
0.61.0-nano.13). A batch ClickHouse rejects for a temporary reason is held and sent again after Vector restarts. A batch it refuses for good is written to a dead-letter folder instead of being dropped, and SIEM Health shows a warning. See Dead-letter folder. - One broken parser no longer keeps ingest closed. If the pipeline can't render because of one parser, nano renders without it, logs it as an error, and opens ingest.
- Faster Risk page. The Risk page API reads much less data, and the page shows each section as soon as its own data arrives.
- Faster case list on installs with many alerts.
- Open in Search from a hunt finding or a case entity now searches the OCSF fields for that entity type (for example source, destination and device IP), not a bare keyword.
- Settings → AI → Agent Models groups the hunt agents under Hunt agents, and Other agents sorts last.
- Case Settings changes are audited in detail. The audit record lists each changed setting with its old and new value.
- No sparkle icons. AI-driven controls carry the pivt mark; other controls get a plain icon for what they do. Case assignee and participant avatars are flat.
Fixes
- pivt no longer changes ingestion settings by itself. Asking pivt to create a log source, add a routing rule, publish a log source or fix a parser now shows a proposal card with each change and the VRL diff. You apply it, and the normal API checks your permissions, validates it and audits it as you. Saving never publishes: publishing is a second, explicit step. pivt never proposes VRL that fails validation.
- Risky AI case actions need confirmation by default. pivt's case actions that close a case, mark it a false positive, lower its severity or unassign it now wait for an analyst's confirmation unless you turn off Confirm pivt case actions in Case Settings. See the upgrade notes.
- Rarity and "newly seen" rules no longer fail on large rescues. Rules using
| prevalencewithhost_countorfirst_seenfilters failed with "Query syntax error at position 262145" when many artifacts were missing from the prevalence dictionary. They now evaluate every artifact.evalandwhereafter| prevalencealso read the prevalence values instead of an empty field. - Root-owned Vector config files are repaired. After some earlier upgrades, files under Vector's dynamic configuration were owned by root, so deploying HTTP ingestion failed with "Permission denied" and quarantine notes couldn't be cleared. Vector now fixes their ownership each time it starts.
scripts/nano upgradelists any that remain. - AI gateway details stay out of logs. nano no longer logs the full AI gateway URL (which includes the account and gateway IDs) or request URLs in AI error messages.
- Storage tiering settings show only on licensed Self-run installs, instead of an error on unlicensed ones.
Upgrade notes
- Required stop at v2.0.20. See above.
- Migrations run automatically. New PostgreSQL and ClickHouse migrations apply during the upgrade.
- Upgrade large installs in a quiet window. One PostgreSQL migration changes a constraint and adds an index on the alerts table, which locks it while it runs. On installs with many alerts, alert writes wait until it finishes.
Ingest during upgrades
-
From v2.0.31 or earlier on a single server, pause direct senders first. The running Vector predates the ingest gate, so the upgrade can't close its ports before it stops Vector. Pause every sender that posts to
8080,8088, OTLP or6000directly. Senders through nginx (/ingest) are closed for you. The installer asks before it fetches or changes anything; on a terminal, typey. -
Unattended runs. With no terminal, the installer refuses with exit code
1and changes nothing unless one of these is set:NANO_CONFIRM_SENDERS_PAUSED=1: you paused the senders;NANO_CONFIRM_UPGRADE=1: answers the upgrade question and confirms the senders too;NANO_CONFIRM_LEGACY_UPGRADE=1: for an open 0.1.x install, confirms the 2.x upgrade and the senders.
scripts/nano upgrade,scripts/nano backupand restore accept the same three. In a terminal, answeringnexits with code3and no answer (end of input) with code4. Nothing is changed in either case. -
From v2.0.32 on, upgrades close and reopen ingest themselves, so this step isn't needed for later upgrades.
Dead-letter folder
- A batch ClickHouse refuses for good is written to
/var/lib/vector/dead_letter/<sink>/inside the Vector container, as a body file plus a.meta.jsonwith the query to replay it. The folder holds up to 256 MiB per sink; beyond that the oldest batches are dropped. - SIEM Health shows a warning with the bytes held and the events stored or dropped per sink. Replay is manual for now: use the replay query in each
.meta.json, then delete the files. - The warning clears the next time Vector restarts (
docker compose restart vector) after you replay and delete the files. - After an unclean Vector stop, a small number of events may be stored twice.
- On role-split ingest servers, SIEM Health doesn't read the dead-letter folders.
AI case action confirmation
- The upgrade turns Confirm pivt case actions on. If you had turned it off in Case Settings, turn it off again after the upgrade. From now on, nano keeps the choice you make there.
Search results can change
- Saved searches, dashboards and rules with staged subsearches can return different results, because the stages inside
IN [ ]now run. A stage that can't run inside a subsearch, areturnthat isn't the last stage, andreturn a, b(a subsearch matches one field) are now refused with an error instead of being ignored. Supported stages inside a subsearch arestats,where,eval,sort,head,dedup,fields,tableandrename. dataset=signalsneeds thedetections:viewpermission on top ofsearch:execute. Detection rules can't read it, as their dataset or in a subsearch.
Hunts
- Hunts need nano Self-run with a valid license, and an AI provider configured in nano. Server hunts don't need Agent Lane turned on. See Hunts administration.
- Case escalation mode needs a track record: the reviewer must be right on at least 90% of at least 20 resolved escalations, judged conservatively. In practice that means about 40 or more. See Escalation.
- Agent rules without a hunt are held. A rule that a person tags
agent-authored, or puts in the Agent detections folder, without an owning hunt is kept paused and archived after 10 minutes. - Runners page. It shows one server row per jobs process. Rows from containers or pods that are gone, including the shared row from before the upgrade, show as offline for up to 7 days, then are deleted. If you disabled the server runner, the disable carries over; to turn it back on, enable every server row with that label.
- Hunts ClickHouse users. If a jobs process dies and no leader is running to clean up, its hunts ClickHouse user can live for up to 4 hours. Its token doesn't count against the hunts limit.
Open 0.1.x installs (legacy upgrade)
- The announced upgrade route is for open 0.1.702 installs on the OCSF 1.8 profile. UDM 0.1.x installs aren't covered: contact nano support. The full runbook is in
docs/administration/upgrades.mdin your install directory, under "Legacy open 0.1.x (OCSF 1.8) to v2". Run the readiness report first. - The installer upgrades a 0.1.x install only on purpose: type
yin a terminal, or setNANO_CONFIRM_LEGACY_UPGRADE=1for an unattended run (it then takes the backup itself unlessNANO_LEGACY_UPGRADE_SKIP_BACKUP=1). - Vector-to-Vector senders on
6000need mutual TLS on 2.x. The installer issues one client certificate bundle per sender intovector-client-bundles/<name>/, with a ready sink config. Name the senders withNANO_VECTOR_NATIVE_SENDERS=edge-1,edge-2, or answer the question in a terminal. See Vector native senders. - OTLP senders on other hosts need action. 2.x publishes the raw OTLP ports
4317and4318on127.0.0.1only. Before the upgrade, either plan to point those exporters at<your nano URL>/otlpwith an ingest token, or setNANO_OTLP_BIND_ADDRESS=0.0.0.0in.envto keep the raw ports open (they stay unauthenticated, so restrict them with your firewall).
Related docs
nano v2.0.33
nano v2.0.33: an nginx security update, run caps and a circuit breaker for SOAR, team and private escalation queues and a duplicate gate for Hunts, one risk score per entity, and search guards based on query cost.
nano v2.0.31
nano v2.0.31: a hotfix. Fresh role-split installs no longer return 503 when a slow ClickHouse health check fails, scripts/nano upgrade no longer fails its SIEM Health check on errors the rollout itself caused, and the installer's re-run upgrade prompt defaults to yes.