Security and data handling
Updated 22 October 2026
/agent-integrations/security-and-data-handling
Security and data handling
The Analyst Agent and the MCP connector read the same TechWolf data. They differ in who runs the AI model and in what TechWolf keeps.
Key attributes
Access follows the Console
Each person needs the TechWolf Agent role in the TechWolf Console, and both agents answer from the data that person can see there, plus TechWolf's Market Intelligence data.
Hosting by geography
TechWolf hosts both agents on Amazon Web Services (AWS), in the geography that holds the organisation's TechWolf data. The Analyst Agent runs only in the EU and the US: organisations whose data is held in Canada can use the MCP connector instead. In Microsoft 365 Copilot, messages first pass through Azure Bot Service and Azure Front Door in TechWolf's own Azure tenant.
No connection to HR systems
Both agents read the organisation's data from TechWolf, and neither connects to the organisation's HR system. Through Microsoft 365 Copilot or another AI tool, TechWolf gets no access to the organisation's files or email: it receives only what a person or the AI tool sends to the agent, such as a question or a search.
Training
Amazon Bedrock does not use the Analyst Agent's prompts or answers to train models. What an AI tool reads through the MCP connector falls under the organisation's agreement with that tool's provider.
Certifications and compliance
TechWolf holds ISO/IEC 27001 and ISO/IEC 42001 certifications and SOC 2 Type 1 and Type 2 reports, complies with the GDPR, and sets out its processing activities in its Data Processing Agreement (DPA). The certificates and reports are available on request from its trust centre at trust.techwolf.ai.
Contacts
Security questions: security@techwolf.ai. Data protection and privacy questions: TechWolf's DPO office at privacy@techwolf.ai.
How it works
| Analyst Agent | MCP connector | |
|---|---|---|
| AI model | Run by TechWolf on Amazon Bedrock, with Anthropic's Claude models. Bedrock may run a request in any AWS region of the organisation's geography, never outside it. | Run by the organisation's AI tool. The connector runs no AI model. |
| What TechWolf stores | Each person's conversations, including the data the agent read and uploaded files, in the agent's region. Each person's name and email address. | Call records, and the registration of each connected AI tool. A sign-in in progress expires after 10 minutes. The employee, job and job family lists fetched for a person stay in memory for up to 5 minutes, for that person only. |
| What is logged | Per question: who asked, when, which tools ran, duration and cost, without the text. The agent's requests to TechWolf's data service are logged too, including search terms. The Amazon Bedrock invocation log keeps the full prompts, including the data the agent read, and the answers. | Per call by a signed-in person: who made it, when, from which AI tool, the tool, its inputs, the outcome and the duration. |
| How long it is kept | Conversations and uploaded files: until the organisation asks TechWolf to delete them. Runtime logs and the Bedrock invocation log: 30 days, in the same geography. The usage record of each question is kept for the organisation's usage reports. | Call records and service traces never hold the data a tool returns. |
| Who at TechWolf can see it | Authorised TechWolf staff, through an internal review tool, and the product team, to improve the agent. | TechWolf, to run and support the connector. TechWolf's internal usage reporting receives the records without their inputs. |
| Employees in answers | Under the identifiers the organisation loaded into TechWolf, with names where it loaded them. | The same, and people can look an employee up by name. |
| Changes to data | Proposed role redesigns, which wait for review, and temporary custom tasks. | None. Every tool only reads. |
| Sign-in | As for the Console, or with a Microsoft account in Copilot. In the Console, the model gets an internal user identifier. In Microsoft 365 Copilot it can get the person's sign-in identifier, which can contain their email address. It never gets a password or key. | In the browser, as for the Console. The AI tool gets a short-lived access token for that person, never a password or API key. |
| Service providers | AWS, including Amazon Bedrock; Auth0, for sign-in; PostHog, in the US, for product analytics without question or answer text; Sentry, for error reports. In Copilot, also Microsoft. | AWS; Auth0, for sign-in; Sentry, for error reports; the AI tool's provider. |
MCP connector
The connector is a remote MCP server over streamable HTTP. AI tools connect with OAuth 2.1, using PKCE and Dynamic Client Registration.
- Consent page. Before sign-in, a TechWolf page names the AI tool as it registered itself and says it wants to read TechWolf data on the person's behalf. People continue only if they started the connection themselves: the name alone does not prove which tool is asking.
- Tokens. An access token lasts about an hour. The AI tool renews it automatically, rechecking the person's access, until about a week passes without use; then the person signs in again. Every call checks the person's Console access, so removing their TechWolf Agent role stops their next call.
- Call records. A call's inputs include a one-sentence reason the AI tool writes, usually the person's question. Service traces also keep the reason. Error reports keep a failed call's inputs and, if it failed unexpectedly, can include a few records it had already read.
- Read-only tools. Each tool reads from a fixed destination inside TechWolf, and no tool takes a URL, accepts a file or sends data elsewhere. An instruction hidden in content the AI tool reads, known as prompt injection, therefore has nothing to change and no way to move data out through the connector. Market answers include text from public job postings written by third parties, so the AI tool's own prompt-injection safeguards apply.
- Third-party tools. The AI tool processes the TechWolf data it retrieves under its own settings, such as memory and data retention.
- Gateways. The connector works behind an MCP gateway that forwards each person's own sign-in, from the same directory as the Console users, not a shared credential. TechWolf checks compatibility with the specific gateway before rollout.
- Network. Each organisation has its own connector address over HTTPS, such as
https://yourcompany.eu3.techwolf.ai/mcp. While connecting, people's browsers open that address, TechWolf's sign-in pages ontechwolf.aiand the organisation's sign-in page. Nothing is installed on devices.
Analyst Agent
- Uploads. People can upload files, such as a list of jobs, in the Console, not in Copilot. Removing a file takes it out of the conversation, but TechWolf keeps it with the conversation.
- Tools. The agent has no web access: its tools reach only TechWolf's data service and a calculation sandbox without network access, so an instruction hidden in an uploaded file or in data cannot make the agent's tools send data out of TechWolf.
- Role redesign. The agent is instructed to send a finished role redesign to the Console only when a person asks, as task and skill changes to one job. Only people allowed to edit and submit job skill and job task changes in the Console can send them, and nothing reaches the job until a reviewer approves the changes and a publisher publishes them.
- Custom tasks. To match a future job profile against existing jobs, the agent can create a temporary custom task for work that has no task yet. It is instructed to remove the task again in the same answer.
Microsoft 365 Copilot
In Microsoft 365 Copilot, TechWolf uses and keeps only the verified email address and the Entra tenant identifier from a person's Microsoft sign-in. The admin consent covers sign-in only: TechWolf reads no directory, mailbox or file content.