Beacon bought Haize Labs to test its own AI, not yours. If Haize red-teams your agents before release or guards them in production, the acquisition announcement does not say who runs that service next quarter. Beacon Software announced on September 17 that it had acquired Haize, with the team becoming Beacon's Applied AI Research Group and co-founder Leonard Tang becoming VP of AI Research, leading the AI operating system deployed across Beacon's 45 portfolio companies. The announcement does not say what happens to Haize's existing enterprise customers. Until someone puts an answer in writing, plan as if the answer is no.
What Did Beacon Actually Buy From Haize Labs?
Beacon bought a team of engineers to point at its own portfolio, and both executives said so in the release. Beacon founder and CEO Nilam Ganenthiran's quote is that "Haize Labs brings exceptional engineers to Beacon who know how to find where AI fails and make it better" — engineers, not a product line. Tang's is that "our work will be put directly to use for Beacon's broad and growing portfolio of essential businesses," according to the full release text. Co-founder Steve Li moves across with him. Terms were not disclosed.
The research group's four stated priorities all face inward: reusable AI infrastructure across the portfolio, domain-specific evaluation and feedback systems, continuous testing, red teaming and observability for AI agents, and AI products that reduce administrative burden and improve customer service, per the release. Beacon calls itself "the first centralized AI holding company" — it buys software businesses serving recreation, utilities, education, government and manufacturing, and runs them on a shared AI platform. That platform is what Tang now leads.
Look at the queue the team is joining. When Beacon closed its $225 million Series C at a $1.4 billion valuation on June 9, it had acquired over 30 businesses. The September release says 45. Every one of them is a customer of the Applied AI Research Group before you are.
The strongest counter-argument comes from Beacon itself. Its homepage describes the company as "a permanent home," says it aims to "preserve the brands and customer relationships that make our businesses valuable," and says the vast majority of its businesses are still run by their founders. If Haize had been bought as the 46th portfolio company, that language would be real reassurance. It was not announced that way. A portfolio company keeps a brand and a customer list; a research group has a roadmap set by its parent. Beacon's homepage does not mention Haize at all.
The same investor sits on both sides. General Catalyst led the August 2024 round that valued Haize at $100 million and co-led Beacon's Series C with HarbourVest Partners. That is common and not improper. It does mean this deal was agreed between two companies with a shared backer, and Haize's customers were not at the table.
What Do Haize Customers Actually Depend On?
Haize sat in three different places in a customer's stack, and each fails differently when a vendor goes quiet. The Next Web describes the business as red teaming, guardrails, evaluation and observability tools that stress-test AI before deployment and monitor it in production — plus red-teaming and safeguards work with the frontier labs, and a pro bono programme for small businesses. A WorkOS profile names the platform's components as Judge, Haize, Monitor and Robustify, and says the company works with OpenAI, Anthropic, AI21 Labs, Google Cloud, Scale AI, HackerOne, Deloitte and MongoDB — a list the profile does not source or split into customers and partners. VentureBeat reported Anthropic as a client in June 2024, roughly six months after the company was incorporated.
AI red teaming is adversarial testing of a model or agent: automatically generating jailbreaks, prompt injections and multi-turn manipulation to find failure modes before users or attackers do. Where it runs decides what breaks:
- The pre-release gate. If Haize scores each build in CI, your release process trusts its pass/fail. When a hosted test service stops responding, a well-built pipeline fails closed and blocks the release. A badly built one treats a timeout as a pass and ships untested. Find out which one you have while the endpoint still answers.
- The runtime guardrail. The same fail-open or fail-closed question, except now it is production traffic, and the wrong default is either an outage or an unguarded agent.
- The assurance record. Findings reports filed as evidence that a system was tested do not break. They age. A red-team report on a model version you replaced in the spring says nothing about the one you run today.
The third is the easiest to miss because nothing alerts. It is also a risk you share with the model providers. The pool of outside firms that test frontier models before release was already small, and the Beacon announcement does not say whether Haize's frontier-lab work continues.
You Can Fork the Judge. You Cannot Fork the Service.
The only Haize asset you can take with you under a clear licence is Verdict, its LLM-as-a-judge library — and it had gone quiet well before the deal. An LLM-as-a-judge is a model used to score another model's outputs against a rubric. Verdict is Haize's MIT-licensed framework for stacking those judges — verification, debate, aggregation — to make the scores more reliable, described in a paper by Nimit Kalra and Leonard Tang. It installs with pip install verdict.
Check the maintenance record before you build on it. The latest release on PyPI is 0.2.7, from July 4, 2025, with a single maintainer listed on a Haize Labs email address. The most recent commit is from November 5, 2025, and it updates a plot. That is more than a year without a release, and the team that maintained it now works on Beacon's platform. The MIT licence means none of that stops you: pin 0.2.7, vendor the source into your own repository, and treat it as code you now own.
Everything else stays behind. The hosted platform, the attack-generation algorithms Haize called its haizing suite, and the red-team engagements are not open source. The public get-haized repository holds jailbreaks discovered by that suite, is described on Haize's own GitHub page as "a subset," and shows no licence — a sample, not a corpus you can build on. The attacks that matter are the ones generated against your system, and those are yours to keep only if your contract says the output belongs to you. Read the data clause before you ask for the export, so you know whether you are asking or demanding.
When Red-Teaming Vendors Sold Before, Who Kept Selling?
Security vendors that bought red-teaming startups kept selling the product under their own name; buyers that wanted the team for themselves made no public promise to customers, and in the two recent cases this publication has tracked, the hosted service was switched off.
| Deal | Announced | What customers got |
|---|---|---|
| Cisco → Robust Intelligence | 2024 | Technology folded into Cisco AI Defense |
| Check Point → Lakera | Sept 16, 2025 | Products renamed Check Point AI Red Teaming and AI Agent Security |
| OpenAI → Promptfoo | Mar 9, 2026 | Written continuity statement; open source stays MIT |
| Harvey → Guardrails AI | Sept 9, 2026 | Hub and hosted inference shut down before the deal |
| Klaviyo → Agency | Aug 5, 2026 | Product wound down Aug 31 |
| Beacon → Haize Labs | Sept 17, 2026 | Nothing announced yet |
Cisco says Robust Intelligence, acquired in October 2024, "has been foundational to the development of Cisco AI Defense." Check Point's announcement made Lakera the foundation of its Global Center of Excellence for AI Security and named the products Check Point AI Red Teaming and Check Point AI Agent Security — Lakera customers got a bigger vendor, not a dead end. OpenAI went furthest. Promptfoo's own announcement says "the team will continue working with customers and users to ensure continuity of service and support," and the repository still says Promptfoo "remains open source and MIT licensed." That sentence is exactly what the Beacon announcement does not contain.
The other pattern is newer. Guardrails AI filed a GitHub issue titled "SUNSET REMOTE INFERENCING AND HUB on 8/25" in July, and Harvey announced it had bought the team two weeks after that cutoff. Agency's homepage says it is being acquired by Klaviyo and is winding down on August 31, 2026 — a notice that, as we reported, never appeared in Klaviyo's own release.
Beacon does not sell AI security. Its companies sell software to — in The Next Web's phrase — everything "from campgrounds to construction companies." There is no external red-teaming product line to fold Haize into, which puts this deal structurally closer to Harvey and Klaviyo than to Cisco and Check Point. That is a pattern, not a prediction. Beacon may yet publish a continuity plan. If it does, get it countersigned: a continuity promise made around a deal is only as durable as the paper it is attached to.
If what you valued was an independent tester, the replacement question is harder than it looks. When Anaconda bought Enkrypt AI, its customers' red-team evidence came to be produced by a company selling an AI platform. Promptfoo, the most familiar open-source fallback, now belongs to a model vendor — weigh that if its reports are evidence about that vendor's models.
What to Do Before Your Next Release Gate
Make your test suite portable before you need it to be — the vendor is replaceable only if the attacks, the scores and the baselines already live somewhere you control.
This Week:
- Inventory every Haize dependency. Search CI configuration, secrets stores and dependency manifests for Haize API keys, SDK imports,
verdictin requirements files, and guardrail endpoints in the request path. Check procurement and expense records for trials a team signed up for on a card. - Ask four questions in writing. Is the platform still sold and supported, and through what date? Which legal entity now holds your contract and your data? How much notice will you get before a shutdown? Will prepaid, unused fees be refunded?
- Set the failure mode deliberately. For every Haize call in a release gate or request path, decide now whether a timeout blocks or passes — then test it by pointing the call at a dead endpoint in staging.
This Month:
- Export the corpus, not just the reports. Attack prompts and full transcripts, findings with severities, judge rubrics and prompts, scores, and baselines per model version. The attacks that broke your system are your regression suite; the tool that found them is not.
- Prove it runs without Haize. Replay the exported attacks through an open-source harness — Promptfoo (MIT), NVIDIA's garak (Apache 2.0) or Microsoft's PyRIT (MIT). If you already gate releases on evals, our regression-gate comparison covers where that replay should live.
- Request deletion once the export is done. Your system prompts, tool schemas and attack transcripts describe exactly how your agents fail. Get written confirmation of deletion — or of which Beacon entity will retain them, and why.
Before Renewal:
- Qualify a replacement against your own corpus. Whether it is open-source tooling run in-house or a commercial vendor such as Patronus AI or Check Point's Lakera-based products, score candidates on how many of your exported attacks they reproduce, not on a demo against somebody else's model.
- Write the clause you did not have. Any red-teaming or eval contract should carry notice of a change of control, a termination right, a data-export obligation in a documented format, and a transition period of continued service after a sale.
- Re-date your assurance file. If a Haize report is evidence in an AI risk register or a customer security questionnaire, record its date and model version, and schedule the retest with whoever replaces it.
The Bottom Line
The independent layer between enterprises and their models is being bought up. In under two years, Robust Intelligence went to Cisco, Lakera to Check Point, Promptfoo to OpenAI, Guardrails AI's founders to Harvey, and now Haize to Beacon. Some of those buyers kept selling; some kept only the people. Either way, the supply of standalone AI testers is shrinking while the number of agents that need testing is not.
The practical answer is the old one for any testing category that consolidates: own the artifacts, rent the machinery. A red-teaming vendor's real output is not a dashboard. It is a growing record of the specific ways your agents fail — and that record is only yours if you hold a copy.
The tool was always replaceable. The record of how your agents break is not. Take it with you.
Continue Reading
- Guardrails' Hub Died Aug 25. Harvey Bought the Team.
- Anaconda Bought Your AI Red Teamer. Read the License.
- Klaviyo Bought Agency. The Product Dies August 31.
- How to Red-Team Your AI Agents Before Production
- Braintrust vs Langfuse vs Promptfoo: Don't Pay for the Gate
- One Firm Ran Three Labs' Cyber Evals. Name Yours.
- NeMo Guardrails vs Guardrails AI vs Lakera: Buy the Detector