Databricks Launches Lakewatch, Buys Two Startups

Databricks entered the SIEM market on March 24 with Lakewatch, a new open, agentic SIEM, and paired the launch with two acquisitions, Antimatter and SiftD.ai. For teams building AI systems in regulated or high-risk environments, the important shift is architectural: Databricks is treating security operations as a lakehouse and agent workflow problem, not as a separate log silo.

Product scope

Lakewatch is in Private Preview. Databricks positions it for defense against agent-driven attacks, with AI agents handling triage, investigation, and response across security, IT, and business data in one governed environment.

Five details define the launch. Lakewatch supports petabyte-scale threat detection and investigations, claims up to 80% lower TCO, handles multimodal data including video and audio, uses open formats, and integrates with Agent Bricks, Genie, Unity Catalog, and Claude models from Anthropic.

If you already run Databricks as a central data plane, this matters because security analytics can move closer to the rest of your governed workloads. The same design pressure shows up across enterprise AI, where agent memory, access control, and tool use increasingly need to live inside one policy boundary.

The acquisitions behind Lakewatch

The two acquisitions explain where Lakewatch is heading.

Antimatter brings security primitives around authentication and authorization for AI agents. Databricks describes the company as having laid the foundation for provably secure authentication and authorization for agents, which aligns with Lakewatch’s emphasis on governed, in-place analysis rather than copying sensitive data into a separate toolchain.

SiftD.ai brings detection engineering and threat analytics depth. Databricks highlights the team’s roots in Splunk’s Search Processing Language and search stack, which is a direct signal that Lakewatch is not just adding model inference on top of logs. It is trying to absorb serious search, detection, and investigation expertise into the product.

This combination is deliberate. One side strengthens policy and control. The other strengthens analyst workflows and detection logic.

Architecture direction

Databricks is pushing a security lakehouse model. Security telemetry, IT records, and business data stay in open formats and are governed with Unity Catalog. Agent Bricks is used to build custom security agents. Genie handles triage and multi-step planning. Claude provides reasoning across the combined data environment.

That matters if your team is already working through AI agent security issues such as prompt injection, over-broad tool access, and weak separation between retrieval and action. Lakewatch’s pitch is that the same platform boundary used for enterprise data governance should also constrain agentic security workflows.

Databricks also includes detection-as-code with automated testing and deployment. For developers, this is one of the more practical product choices in the launch. Detection logic becomes something you can version, test, and ship, closer to modern software delivery and closer to how teams already think about evaluating agents.

Cost and retention strategy

The economic argument is central, not secondary. Databricks says traditional SIEM pricing forces teams to discard up to 75% of their data. Lakewatch is designed around retaining and analyzing years of data without moving or duplicating it, with a stated goal of up to 80% lower TCO.

Those numbers matter because they point to a different operating model.

Capability	Lakewatch
Availability	Private Preview
Positioning	Open, agentic SIEM
Data scale	Petabyte-scale detection and investigations
Data types	Security, IT, business, plus video and audio
Governance	Unity Catalog
Agent layer	Agent Bricks, Genie
Model integration	Claude models from Anthropic
Cost claim	Up to 80% lower TCO

If your SOC currently samples or drops telemetry for cost reasons, Databricks is offering a way to keep more raw history and let agents work over a larger evidence base. This follows the same pattern seen in long-context and retrieval-heavy AI systems, where keeping more context changes what the system can reliably infer. The tradeoff is operational discipline, because more retained data only helps if your policies, tool permissions, and context construction are solid. Work on context engineering becomes part of security engineering.

Ecosystem position

Databricks launched Lakewatch with a large partner set, including Akamai, Okta, Palo Alto Networks, Panther, Proofpoint, Slack, Wiz, and Zscaler. Adobe and Dropbox are named as customers.

The partner list matters because SIEM products live or die on ingestion, normalization, and workflow coverage. Databricks is using ecosystem breadth to reduce the usual objection to new security platforms, which is that teams cannot afford to rebuild every integration path.

Anthropic’s role also matters. Claude is being used for reasoning across mixed enterprise data, and the partnership gives Databricks a credible model layer for investigations that need strong long-context synthesis. If you are comparing model stacks for enterprise agents, the choice still depends on your workflow and governance requirements, similar to broader tradeoffs in GPT vs Claude vs Gemini.

Databricks is betting that security operations will converge with governed data infrastructure and agent tooling. If you already centralize enterprise data in Databricks, the practical next step is to evaluate whether your detection logic, access controls, and investigation workflows can run inside the same platform boundary, because Lakewatch is designed around that assumption.