SOC 2

SOC 2 is a reporting framework from the American Institute of Certified Public Accountants (AICPA). An independent audit firm examines a service organisation's controls against the AICPA's Trust Services Criteria, of which security is the foundation; availability, confidentiality, processing integrity and privacy can be added to the scope.

There are two types of report. A Type I report considers whether the controls are suitably designed at a single date. A Type II report goes further and tests whether they operated effectively over a period, usually between three and twelve months, which is why it is the report most firms ask for. SOC 2 does not result in a certificate. The outcome is the auditor's report, which describes the controls, the tests performed and any exceptions found, and which is shared with clients under a confidentiality agreement.

ISO/IEC 42001

ISO/IEC 42001, published by ISO and the IEC in December 2023, is the first international standard for an AI management system. It follows the same structure as ISO/IEC 27001 and, like it, can be certified by an accredited body. Where SOC 2 asks whether a service and its data are protected, ISO/IEC 42001 asks whether the organisation governs the AI itself: who is accountable for it, how the impact of each AI system on people is assessed before it is used, how people oversee it, and how it is monitored and improved over its lifecycle.

How the two fit together

The two standards overlap in places, such as access control, supplier management and incident handling, but neither replaces the other. A supplier can hold a clean SOC 2 report and still have no defined process for deciding whether an AI system should be used at all. Equally, a well-governed AI programme says little about whether the infrastructure beneath it is secure. For that reason a firm placing client data with an AI supplier may reasonably ask to see evidence of both.

Where Lynctera stands

Our first SOC 2 Type II audit is complete, covering the period from August to October 2025. Our second observation period runs from November 2025 to the end of October 2026, and the report will follow once the auditor has completed its review. The current report is available on request through the Yellow Be trust centre.

Our ISO/IEC 42001 programme is in progress. We are not certified, and we have not set a public date for certification. Several of the controls the standard asks for are already part of how the platform works. Each AI agent we run has an impact assessment record, generated from its live configuration and kept current by our release checks, and completing the written risk analysis for every agent is part of the work still in progress. A named person signs off before anything consequential happens, and every decision is recorded. Anyone can report an adverse impact of our AI through Yellow Be's reporting form. We will post here when the programme reaches certification.

Once ISO/IEC 42001 is in place, we intend to pursue ISO/IEC 27001, the international standard for information security management. SOC 2 already covers much of the same ground, and because the two ISO standards share a structure, much of the management system built for one serves the other. As with ISO/IEC 42001, we have not set a public date.

A fuller description of the standard is on our ISO/IEC 42001 page, and the controls that apply across the platform are described on Governance.

If your risk or compliance team would like to go through either assurance in more detail, we would be glad to arrange it.

Using Lynctera from Claude

For people who already use Claude in their daily work, the Lynctera connector makes the platform available from there. They can see what has changed on their cases, open a case, start a research run, approve a step that is awaiting them, or have a reply drafted in their own voice and saved to Outlook.

Claude directs the work, while the handling of confidential material stays within Lynctera. Our agents work inside each client's own isolated workspace, and Claude receives the outcome rather than the archive: a status, a draft, or a link to the relevant step in the portal. Email bodies, the passages of a precedent library and the full text of documents are not available to Claude under any setting, and the connector has no means of sending anything.

Every request carries the person's own Microsoft sign-in and is limited to what that person could see in the portal. Each answer is checked before it leaves Lynctera, so that text written by someone outside the firm is withheld, and every request is written to the firm's own audit record.

The firm's decision

Whether anything reaches Claude is a decision for the firm. The connector is switched off by default, and a firm that wishes to use it switches it on one area at a time, from a consent page that sets out what each area shares and where it goes. Before doing so, the firm confirms that its people use Claude through an organisation account and that it has considered its confidentiality obligations to its own clients. Drafting and cases can be switched on today, and precedents, email and market data will follow.

Every change is recorded. The firm can see who has connected and disconnect anyone, and people can see and end their own connections from their Lynctera settings. Switching an area off stops new requests at once, although it cannot recall information already returned to Claude.

Anything Claude receives, and anything typed into Claude, is held by Anthropic under the firm's own agreement with it, outside our platform and outside the regions in which we host the firm's data. For that reason the connector sends no more than each task requires.

The connector is in use on our own workspace. Client workspaces will be able to connect once sign-in from the client's own Microsoft directory is in place, and we will post here when they can.

Outside services used by our agents

Our agents also use MCP to draw on outside data, such as market prices and earnings call transcripts. Because the operator of a service can change its tools at any time, we approve each tool individually and record its exact definition. If the vendor changes it, the tool is switched off automatically until a person has reviewed the change. Every tool connected today is read-only, and whatever it returns is treated as outside content rather than as instructions.

Governance

MCP does not change the rules that govern the rest of the platform. Approvals given through Claude are the person's own, actions taken through Claude enter the same decision record as those taken in the portal, and each agent's AI impact assessment lists the outside tools it holds. These rules are described on Governance.

If your risk or compliance team would like to go through the detail, we would be glad to arrange it.

Where Claude is used today

Claude Opus 5 writes our longest documents, such as a sector note drafted from two years of research. Claude Sonnet 5 reviews new standing instructions for risk before they take effect, and builds the writing-style profile that allows drafts to read like the person sending them.

Before any agent moves

Each of those agents will be replayed on the 5.5 model against its own checks, and a person will review the comparison. An agent moves only if the new model performs at least as well on that agent's work. This is how every model change on the platform is made, as described on Models.

For Opus, the price alone makes the comparison worthwhile. At list price, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5. Sonnet 5.5 is priced the same as Sonnet 5, at $2 and $10.

We will report the results here once the comparisons are complete.

The comparison

GPT-6 Luna reached our production platform on 23 September. That day we replayed real work from our own mailbox through three of the agents, once on the model they were then using, GPT-5.6 Luna, and once on GPT-6 Luna, and scored every output against each agent's own checks.

Times are the average per answer. At list prices, each answer cost between 38% and 49% less on GPT-6 Luna, even though the model writes more tokens because it reasons before answering.

An issue found on the first run

On the first run, three of the 28 summaries came back empty. GPT-6 Luna spends more of its output allowance on reasoning, and in those three cases the allowance ran out before any summary was written. We raised the allowance from 1,500 to 4,000 tokens and ran the comparison again: all 28 summaries were complete, with no loss of speed. The higher allowance went to production together with the model.

What remains open

The reply-draft sample was small, at ten drafts. On it, GPT-6 Luna suggested a next step in eight drafts, against nine for GPT-5.6 Luna. We are monitoring this in production rather than treating ten drafts as conclusive either way.

The models our agents use, and how the work is divided between them, are listed on Models.