Safeguard
Product•

Organize AI-assisted security work: workspaces, skills, connectors

Set up shared Griffin AI Search workspaces, reusable skills with MCP tool whitelists and search connectors in Safeguard so security investigations repeat across a team.

Safeguard Security Team
36 min read

To make AI-assisted security investigations repeatable in Safeguard, create a Workspace in Griffin AI Search that holds a shared system prompt, a default model and a member list, compose one or more Skills into it (each Skill is a reusable prompt with an optional whitelist of MCP tools), and connect Search Connectors so Griffin, Safeguard's security model, can call outside tools under per-tool permissions. Every conversation started in that workspace then begins with the same framing, the same tool boundary and the same people.

Below: each building block as the Safeguard docs describe it, which layer wins when they disagree, and a step by step setup for a shared triage workspace.

Key takeaways

  • A Workspace groups Search conversations under a shared system prompt, a default model and a member list, "like a project". Conversations started in it inherit that context automatically.
  • A Skill is a reusable system prompt you invoke with / in the Search composer. It can carry a whitelist of exact MCP tool names, which restricts the tool registry for that turn to those tools only.
  • Skills are private by default. Marking one Public shares it across your tenant so everyone can invoke the same triage logic.
  • A Workspace can compose several Skills, so every conversation in it runs with the workspace prompt plus each Skill's prompt layered on top, without anyone remembering to type /.
  • Search Connectors attach external MCP servers as live tools Griffin can call mid-conversation. Each tool has a permission: Always allow, Needs approval (the default) or Deny.
  • Precedence is predictable: a Skill applied to the chat, or a setting on the conversation itself, overrides the workspace's system prompt and default model.
  • These features organize how people use AI inside Safeguard. They sit next to, and do not replace, tool gating on Safeguard's own MCP server and Guard for MCP servers and agents you run yourself.

The three building blocks at a glance

Before going deep, here is how the three features divide the work. Each one answers a different question a team lead asks when AI-assisted work starts to spread.

Building blockQuestion it answersWhere you manage itHow you invoke it in a chatWho can use it
Workspace"What context should every conversation about this project start with?"Workspaces tab (spans products)Type > in the composerOwner and members
Skill"How exactly should this recurring task be done, and with which tools?"Integrations, then SkillsType / in the composer, or compose into a workspaceCreator, or the whole tenant if Public
Search Connector"Which outside systems may Griffin call during a conversation, and how freely?"Connector dialog in the Search interfaceGriffin calls connector tools as needed; & references a connectorGoverned by per-tool permissions

The composer also takes @ for a project, # for an SCM configuration and : for a specific entity such as a component or vulnerability, so an analyst can point at exact scope instead of describing it.

Why repeatable AI-assisted investigations need structure

AI assistance arrives in a security team one person at a time. Someone writes a good CVE triage prompt; it lives in their chat history. A colleague asks the same question differently a week later and gets a differently shaped answer, using different criteria and different tools. The team cannot say how a conclusion was reached, and two analysts can reach different answers from the same data. Public guidance explains why that drift matters.

Excessive agency. The OWASP Top 10 for LLM Applications names three root causes of Excessive Agency: excessive functionality, excessive permissions and excessive autonomy. Its mitigations include limiting extensions to the minimum needed and requiring human approval for high-impact actions. A Skill's tool whitelist and a connector's per-tool permissions express exactly those two mitigations.

System prompts are not a security boundary. OWASP's System Prompt Leakage entry says the system prompt should not be considered a secret nor used as a security control, and credentials should not be placed in it. Workspace and Skill prompts are for consistent method; access comes from roles, memberships and tool permissions.

Human in the loop for tools. The Model Context Protocol specification says that for trust and safety there should always be a human in the loop with the ability to deny tool invocations, and that applications should make clear which tools are exposed to the model and show clear indicators when tools are invoked. It also says clients must treat tool annotations as untrusted unless they come from trusted servers. Search Connectors default every tool to Needs approval, which is the conservative reading of that guidance.

Wider frameworks point the same way. The NIST AI Risk Management Framework (released 26 January 2023, voluntary), its generative AI profile, NIST AI 600-1, and the UK NCSC's Guidelines for secure AI system development all treat governance, documented method and managed risk as part of using AI safely. One habit follows from them: decide the method once, write it down where everyone uses it, constrain what the system may do, and keep a human on the actions that matter. Workspaces, Skills and Search Connectors are where that habit lives inside Safeguard.

Workspaces: one context for a team's conversations

The Safeguard docs describe a Workspace as a conversation grouping in Search that bundles a system prompt, a default model, composed Skills and a member list, "so every conversation you start inside it shares the same context and configuration instead of you re-explaining it each time." The dialog that creates one sums it up in a line: "Group conversations under a shared system prompt and default model, like a project."

What a Workspace holds

FieldWhat it doesLimit or default
NameIdentifies the workspace, for example "Q1 Security Review"Required, up to 80 characters
DescriptionOne-line summary shown in the workspace listOptional, up to 280 characters
System promptInstructions prepended to every conversation started in the workspaceOptional, up to 8,000 characters
Default modelModel applied to conversations in the workspace unless overriddenOptional
ColourAccent colour (hex, for example #6366f1) shown in the sidebar and detail headerOptional
SkillsSkills composed into every chat in the workspaceZero or more
MembersThe users who have access to the workspaceOwner plus members
EnabledDisable a workspace without deleting itToggle

Two details matter. The Create panel sets only name, description and system prompt; default model, colour and Skills are added afterward from the Edit tab. And a conversation belongs to at most one workspace at a time. When it does, the workspace's system prompt and default model apply automatically, unless a Skill applied to that chat or a setting on the conversation itself overrides them.

Why a Workspace beats a pinned prompt in a wiki

The docs' reasons map to failure modes teams already know.

  • Consistent framing. A system prompt tuned for one kind of question, such as incident response or an audit, applies to every conversation in the workspace, not only the one where someone typed it.
  • Consistent model. You set the default model once instead of each analyst picking one per conversation.
  • Shared access. Anyone added as a member can use the workspace and its conversations, so context is not trapped in one person's chat history.

Shared access is the one leads underrate: whoever picks up an investigation after a shift change opens the same conversations with the same framing.

Managing workspaces

Workspaces span products, so the Workspaces tab shows the same list whichever product you are in. The table shows each workspace's name, description, default model, member count and last-updated time, with an enabled toggle and row actions to open details or delete.

Opening a workspace shows two tabs:

  • Overview is a read-only summary: default model, member count, colour, created and updated timestamps, the full system prompt and the list of composed Skills.
  • Edit is where you change the name, description, system prompt, default model, colour (with a swatch picker) and which of your Skills are composed into the workspace. Save changes applies them.

Attaching a workspace to a conversation

From the Search composer, type > to open the workspace picker. It lists your workspaces with name and description and filters as you type, so >Q1 narrows the list to workspaces whose name matches "Q1". Select one to attach it to the current conversation, or choose Create new workspace to open the create panel without leaving the composer. The > picker mirrors the / Skill picker: / for how, > for where.

Member access

Each workspace has an owner and a list of members. Members get shared access to the workspace's system prompt, default model and composed Skills, the same way the owner does. The Workspaces table and the Overview tab show this as a member count badge, for example "3 members", rather than listing individuals inline. For a security lead that means membership is the unit of access to a team's AI context, and it should be reviewed like any other access list.

Skills: a reusable method with a tool boundary

The docs define a Skill as "a reusable system prompt, optionally paired with a whitelist of MCP tools, that you trigger by typing /skill-name" in the Search input. When a Skill is invoked, the model call for that turn is prefixed with the Skill's system prompt. If the Skill has a tool whitelist, the tool registry for that turn is restricted to only those tools.

That second half turns a saved prompt into a control: a prompt can ask the model to stay within certain tools, while a whitelist makes everything else unavailable for that turn.

What a Skill is made of

FieldDescriptionLimit or default
NameThe identifier you type after /Up to 80 characters
DescriptionShown beside the name in the management table and the / pickerOptional, up to 280 characters
System promptText prepended to the model call for the turn whenever the Skill is invokedRequired, up to 8,000 characters
Allowed MCP toolsComma-separated list of exact tool names; restricts the turn's tool registry to these toolsOptional, up to 2,000 characters; empty means no extra restriction
PublicShares the Skill tenant-wideOff by default
EnabledA disabled Skill is skipped at compose timeOn by default

The docs give a worked example worth copying as a starting point:

Name:              Vulnerability Triage
System prompt:     You are a vulnerability triage assistant. Respond in
                   formal regulatory tone, cite frameworks.
Allowed MCP tools: safeguard_find_vulnerabilities, safeguard_list_projects
Public:            off for your own workflow, on if the whole team
                   should triage the same way

With that saved, typing / in Search and selecting Vulnerability Triage inserts the Skill as a token in your message and attaches its system prompt to the turn, with the tool registry restricted to those two tools "no matter what else is available."

Creating and managing Skills

Skills live on the Integrations page:

  1. Go to Integrations, then Skills.
  2. Click New skill and fill in a name, an optional description, the required system prompt, and optionally a tool whitelist and the Public flag.
  3. Save. The Skill appears in your Skills table and is available as /name in Search.

Clicking a row, or the eye icon, opens the Skill's detail panel inline. It has the same two-tab pattern as workspaces:

  • Overview shows visibility (Public or Private), how many tools are whitelisted (or "Unrestricted" if none), timestamps, the full system prompt, the whitelist as badges, and which conversations have used this Skill. That last item is useful evidence when you review whether a Skill is actually adopted.
  • Edit carries the same fields as the create dialog, saved with Save changes.

New conversation in the detail panel starts a chat with the Skill attached. From the table you can toggle Enabled or Disabled (effective immediately) or delete a Skill, which asks you to confirm by name and cannot be undone.

Invoking a Skill with "/"

Type / in the Search input to open the picker: your own Skills plus public Skills in your tenant, filtered as you type. Others' public Skills carry a Public label. Create new skill opens the create dialog in place.

Public versus private Skills

Skills are user-scoped by default, visible only to their creator. Setting Public shares a Skill tenant-wide in everyone's picker. Sharing does not move ownership. That is why public Skills created by others show up in your / picker but not in your own Skills management table, which lists only Skills you created.

Adopt the docs' guidance as policy: use a private Skill for something personal to your own workflow, and a public Skill for something you want the whole team invoking consistently, "like a shared CVE-triage prompt everyone should be running the same way."

The whitelist takes exact names

The Allowed MCP tools field is a plain comma-separated list. You type each tool's exact registered name. A typo or near-miss name is simply excluded from the whitelist rather than flagged as an error. Two consequences follow:

  • If you whitelist a tool and the model still cannot use it, open the Edit tab and check the spelling of each entry against the registered names.
  • If the whitelist is empty, no extra restriction applies for that Skill, and the turn's tool registry is unrestricted. If the model used a tool you expected it to avoid, the first thing to check is whether the list is empty.

Treat the whitelist like a firewall rule set: write it, then test it by asking the Skill to do something just outside its scope and confirming it cannot.

Search Connectors: live tools from outside Safeguard

Search Connectors let Griffin reach outside your Safeguard data during a conversation by calling tools on external MCP servers you have connected. The docs are careful to separate this from the Connectors catalog, which pulls one-time or scheduled compliance evidence from cloud, identity, HR and other systems into the compliance pipeline. A search connector is not an evidence pull. It is "a live tool source Griffin can invoke mid-conversation, the same way it would call any other tool."

You manage search connectors from the connector dialog in the Search interface. It has two tabs: Browse Directory and Custom MCP.

Browse Directory

Browse Directory, the default tab, is a searchable grid of MCP servers from Safeguard's connector directory; typing "slack" narrows it to Slack-related entries. Each entry shows an auth badge (OAuth, API key, Basic, or Public for no credentials) and a Featured badge for curated entries. Entries with a ready-to-use MCP endpoint sort to the top.

Connect on an entry opens a panel that offers up to three methods together, so you pick whichever fits:

MethodWhat it doesWhen to choose it
MCP server (recommended)Uses the entry's known MCP endpoint, pre-filled and marked Ready. You can add an access token sent as a bearer credential. With the token blank, Safeguard also attempts a background OAuth handshake; if the server does not support it, the connector is added for you to authorize later from its detail pageDefault for any entry with a known endpoint
OAuth 2.0Signs in with the provider's own OAuth appWhen an admin has configured that provider's OAuth client in Safeguard
Personal access tokenPaste a token. With a known MCP URL it is sent as a Bearer credential; otherwise it is stored as the entry's API key credentialProviders without OAuth, or service accounts

For a GitHub-style entry with a known endpoint, the docs' flow is: choose MCP server, leave the token blank, click Connect via MCP, and approve the OAuth tab that opens. If the popup is blocked or closed early, retry from the dialog or authorize from the connector's detail page.

The message "OAuth needs a one-time admin setup for this provider" means an admin must configure that provider's OAuth client in Safeguard once; use another method meanwhile.

Registering a custom MCP server

For a server that is not in the directory, such as an internal documentation server or a runbook service your platform team runs, use Custom MCP:

FieldRequiredNotes
NameYesLabel for the connector, up to 80 characters
Server URLYesThe Streamable HTTP or SSE endpoint of your MCP server; must be a full http:// or https:// URL
Authorization headerNoSent as the Authorization header on every request, up to 2,000 characters, for example a bearer token your server expects

Submitting registers the server so Griffin can start calling its tools. Test it in a conversation straight after, with permissions set conservatively.

For the team that runs the server: the MCP specification's Streamable HTTP transport replaces the older HTTP plus SSE transport from protocol version 2024-11-05, and its security warning says servers must validate the Origin header and should implement proper authentication. A server reachable by Safeguard is reachable over the network, so treat authentication as mandatory.

Tool permissions: the control that matters most

Once a connector is connected, open it from the connectors list to see what it can do. Its tools are grouped into three categories by what each tool does: Interactive tools, Read-only tools and Write/delete tools. Each tool, and each category as a whole, has a permission:

PermissionEffect
Always allowGriffin can call the tool without prompting you
Needs approval (default)Griffin asks you to confirm before each call
DenyThe tool is never callable

Setting a category's permission applies it to every tool currently in that category, and you can still override an individual tool afterward. A safe starting pattern: set the whole Write/delete category to Deny, set Read-only to Always allow only for tools you have reviewed, and leave Interactive at Needs approval. Approval matters for reads too: OWASP lists indirect prompt injection, instructions hidden in content a tool returns, as a leading LLM risk, so outside data deserves a human glance before the model acts on it.

If a connector's tools fail to load, for example because a stored credential has expired, the detail page shows "Could not load tools" in place of the list. Fix the credential, then use Refresh tools from the connector's menu.

Disconnect, remove and credential storage

Disconnect on a connector's detail page keeps it in your list; only then does its menu offer Remove, which deletes it.

The docs say any token you paste for a directory entry's MCP server or personal access token method, and any authorization header you set for a custom server, are encrypted at rest server-side with Fernet encryption before storage. The Fernet specification describes symmetric encryption with AES 128 in CBC mode and a SHA256 HMAC. Safeguard does not display stored credentials back to you; a connected entry shows only as connected.

Because you cannot read a stored token back, rotation means issuing a new token at the provider and replacing it in Safeguard. For a connector added with a personal access token, the docs' fix for an expired or revoked credential is to remove it and re-add it with a fresh token, then Refresh tools to confirm.

RFC 6750 reminds us that "any party in possession" of a bearer token can use it. Give each connector the narrowest read-scoped token the provider offers, or use OAuth so there is no long-lived token to paste. The MCP authorization specification builds on the OAuth 2.1 draft and requires clients to use Resource Indicators (RFC 8707) so a token is requested for a specific MCP server, and its security best practices page calls out "token passthrough", where a server accepts tokens not issued to it and forwards them downstream, as an anti-pattern.

How the layers combine, and which one wins

The three building blocks stack. Knowing the order of precedence prevents the most common support question: "why did the workspace prompt not apply?"

SituationWhat applies
Conversation in a workspace, no Skill invoked, no conversation overrideWorkspace system prompt and default model, plus every Skill composed into the workspace
Conversation in a workspace, a Skill applied to the chat whose instructions conflict with the workspaceThe Skill takes precedence for that conversation
Conversation in a workspace with its own setting overrideThe conversation's setting overrides the workspace's system prompt and default model
A Skill with a tool whitelist is invokedThe turn's tool registry is restricted to the whitelisted tools
A Skill with an empty whitelist is invokedNo extra tool restriction for that turn
A Skill is disabledIt is skipped at compose time; invoking it has no effect on prompt or tool access
A connector tool is set to DenyIt is never callable, whatever any prompt says
A connector tool is set to Needs approvalGriffin asks before each call

Read it this way: prompts shape behaviour and permissions bound it. A workspace prompt can say "never modify anything", but the reliable controls are a Skill whitelist that leaves out write tools and Deny on a connector's Write/delete category. Composition is additive (the docs' example composes a soc2-controls and a cite-sources Skill into one workspace), while an override in one chat changes that chat only.

Step by step: build a shared triage workspace

The case: a product security team triages new vulnerabilities across its estate and wants every analyst to work the same way, using Safeguard data plus a ticketing or chat system connected as a search connector.

Step 1: write the method down first

Write the triage method as plain text first. It becomes the Skill prompt, and writing it outside the product forces the team to agree on it:

You are assisting a product security analyst with vulnerability triage.

For each vulnerability in scope:
1. State the CVE ID, severity, affected package and version, and the fixed
   version if one is known.
2. List the projects that contain the affected version.
3. Say whether the vulnerability is marked reachable and whether an exploit is
   known, and name the data you used.
4. Recommend one of: fix now, fix this sprint, accept with mitigation,
   not affected. Give one sentence of reasoning.
5. If data is missing, say what is missing. Do not guess versions.

Output a table first, then a short summary. Cite the Safeguard record for
every claim.

Keep credentials and anything you would regret seeing quoted back out of it, per OWASP's System Prompt Leakage guidance.

Step 2: create the triage Skill with a tight whitelist

  1. Open Integrations, then Skills, then New skill.
  2. Name: triage-cve. Description: House method for triaging a CVE across our projects.
  3. System prompt: paste the method from step 1.
  4. Allowed MCP tools: start from the docs' example, safeguard_find_vulnerabilities, safeguard_list_projects, and add only the other read tools your method needs, typed exactly as registered.
  5. Leave Public off for now.
  6. Save, then open the detail panel and check the Overview tab: the whitelist shows as badges, and the count should match what you meant to allow.

Step 3: test the Skill before you share it

From the Skill's detail panel, click New conversation. Run two tests:

  • A normal case. Ask it to triage a vulnerability you know is in scope. Check the output shape matches the method.
  • A boundary case. Ask for something that would need a tool you deliberately left off the whitelist. The Skill should not be able to do it. If it can, the whitelist is empty or misspelled.

When both pass, turn Public on so the whole tenant sees triage-cve in the / picker, labelled Public.

Step 4: create the workspace

  1. Open the Workspaces tab and click New workspace (or type > in the composer and choose Create new workspace).
  2. Name: Product Security Triage. Description: Shared triage and weekly exposure reporting.
  3. System prompt, kept short because the Skills carry the method:
This workspace is used by the product security team for vulnerability
triage across our products. Prefer Safeguard data over general knowledge.
When a fact comes from a connected outside system, say which one.
Ask before any action that changes something outside Safeguard.
  1. Click Create workspace.

Step 5: finish configuring from the Edit tab

Open the new workspace, go to Edit, and:

  • Set the default model your team has agreed for triage, so nobody picks one per chat.
  • Pick a colour so the workspace is easy to spot in the sidebar.
  • Check triage-cve (and any reporting Skill you build the same way) under Skills, so it composes into every conversation in the workspace.
  • Click Save changes, then check the Overview tab shows the system prompt and the composed Skills.

Step 6: add members

Add the analysts who run triage as members, and only them; the member count badge is your quick check. Shared context deserves the same access hygiene as a shared drive.

Step 7: connect outside tools with conservative permissions

Open the connector dialog in the Search interface:

  1. In Browse Directory, search for the system you use for tickets or team chat. Choose MCP server if the entry has a Ready endpoint, leave the token blank to use OAuth if the provider supports it, and approve the OAuth tab.
  2. If your platform team runs an internal MCP server (for example, a runbook lookup service), add it under Custom MCP with its full https:// URL and an authorization header holding a token scoped to read only.
  3. Open each connector and set permissions by category: Write/delete tools to Deny, Read-only tools to Needs approval at first, Interactive tools to Needs approval.
  4. After a week of use, move individual read tools you have reviewed to Always allow. Leave everything else asking.

Step 8: run the first investigation

In Search, type >Product and select the workspace. Then use the composer's mentions to scope the question precisely: @ to pick the project, : to pick the vulnerability entity, and & to reference the connector if you want Griffin to draft a ticket comment for your review. Because triage-cve is composed into the workspace, you do not need to type it. Star the conversation if it is a sweep you will repeat, and it appears under Starred searches in the sidebar.

Step 9: review and adjust

After a week, open each Skill's Overview tab to see which conversations used it. An unused Skill has a wrong method or an unclear name; workarounds mean the whitelist is too tight. Change the Skill once and every member's next conversation picks it up.

Writing prompts that hold up across a team

State the method, not the persona. "You are a world-class security analyst" adds nothing a reviewer can check. "List the projects that contain the affected version, then say whether it is marked reachable" is a step someone can verify in the output.

Define the output shape. Tables with fixed columns make answers comparable across analysts and across weeks. If the output feeds a ticket, mirror the ticket fields.

Say what to do when data is missing. "If the data is not there, say so; do not guess versions" turns a confident wrong answer into an honest gap.

Keep workspace prompts short. Framing goes in the workspace; methods go in Skills, reusable and changed in one place. Prompts can run to 8,000 characters, but long prompts are hard to review.

Never put secrets or access logic in a prompt. Credentials belong in a connector's encrypted credential store. Who may see what belongs in workspace membership, Safeguard roles and tool permissions.

Version prompts outside the product too. Keep the text of each public Skill in your team's repository next to the runbook it implements. When someone changes the Skill, the diff is reviewable, and if a Skill is deleted by mistake (deletion is permanent) you can recreate it exactly from the file.

Designing tool whitelists and connector permissions

Classify each workflow by what it must be able to do.

Workflow typeSkill whitelistConnector permissionsWhy
Exploratory question ("what is our exposure to this package?")Read tools for projects, vulnerabilitiesRead-only: Needs approval, then Always allow once reviewed; Write/delete: DenyWide reading, no changes
Standard triageThe exact read tools the method usesSame as aboveKeeps every analyst on one tool path
Drafting tickets or messagesRead tools onlyThe connector's drafting tool: Needs approval, so a person confirms each callA human approves each outward action
Executive or audit summaryRead tools onlyAll connector tools Deny unless the summary needs outside dataOutput should rest on Safeguard records
Anything that deletes or modifies outside systemsDo not build it as a Skill until you have a reviewed needWrite/delete: DenyExcessive agency risk outweighs convenience

Two MCP specification rules make good team policy. Keep a human able to deny tool calls: move a tool to Always allow only after reading what it does. And do not trust a tool's description of itself, since the spec says annotations are untrusted unless the server is trusted; a tool named like a read is not proof that it only reads.

Governance: ownership, review and change control

Without a little ownership, shared work decays back into personal prompts.

  • Name an owner for every public Skill. The creator owns it, and sharing does not move ownership. If the creator changes teams, recreate the Skill under the new owner from the version in your repository.
  • Use a naming convention. Prefix public Skills by team (appsec-triage-cve) so the picker stays readable.
  • Prefer disable over delete. A disabled Skill is skipped at compose time and can be re-enabled. Deletion is permanent. The same applies to workspaces, which have an Enabled toggle.
  • Review membership and connectors on a cadence. Once a quarter, check each workspace's member list, each connector's permissions, and each public Skill's whitelist. Disconnect connectors nobody uses, and remove them after a waiting period.
  • Rotate connector credentials deliberately. Stored credentials are never shown back, so rotation is replace, not inspect. When a provider token expires, expect "Could not load tools", and follow the reconnect or remove and re-add path the docs describe.

Where this sits next to Safeguard's MCP server, tool gating and Guard

Search Connectors connect Griffin outward. Safeguard's MCP server runs "the reverse direction", in the docs' words: an outside assistant such as Claude, ChatGPT, Cursor, VS Code or Gemini CLI connects inward to your Safeguard data. Three controls govern that direction, none set by a workspace or Skill:

  • OAuth sign-in or an MCP key. OAuth is recommended for clients that support it. An MCP key, for a client that can only send a fixed header, acts as the member who created it across your whole Safeguard tenant (not just one Search Workspace), is shown once, and stops working within a minute of being revoked. Our MCP key safety checklist for Safeguard workspaces covers handling and revocation in detail.
  • Tool availability and gating. Administrators decide per tool, under Feature Flags, which of Safeguard's own MCP tools an outside assistant can see and call. That is separate from a Skill's whitelist, which governs a turn inside Safeguard Search.
  • Guard. The docs describe Guard as a separate product that enforces live policy on MCP tool calls in general, including MCP servers and agents you build or operate, for example by running Guard Proxy in front of servers whose code you do not control. Tool gating and Guard are independent: changing one does not change the other.

If you are designing permissions for agents beyond Safeguard Search, our posts on tool permission design for agents and AI coding agent governance go further into scoping and audit trails. And if your organization needs AI features off, or wants its own model, read Zero AI mode and bring your own model for how the platform runs in each mode.

Common mistakes

Putting the method in the workspace prompt. It works for one workspace, then the method drifts as people copy it into others. Put the method in a Skill, compose the Skill into each workspace that needs it, and change it in one place.

Leaving the whitelist empty and assuming the prompt restricts tools. An empty whitelist means no extra restriction. Only the whitelist restricts the turn's tool registry.

Misspelling a tool name. The whitelist does not flag typos; a near-miss name is silently dropped. Check the Overview tab's badge list after every edit.

Expecting composed settings to override everything. A Skill applied to a chat, or a conversation's own setting, wins over the workspace's system prompt and default model. When a workspace prompt "does not apply", check for those first.

Switching whole connector categories to Always allow. It is quick and it removes the human from every call in the category. Approve per tool after review.

FAQ

What is a workspace in Safeguard Search?

A Workspace groups Search conversations under a shared system prompt, default model, composed Skills and member list, like a project. Conversations started in it inherit that configuration.

How do I create a reusable prompt for my security team?

Create a Skill under Integrations, then Skills: give it a name, a required system prompt and, optionally, a whitelist of exact MCP tool names. Turn on Public to share it tenant-wide. Teammates invoke it with / in the Search composer, or you compose it into a workspace so it applies automatically.

Can I restrict which tools the AI can use during an investigation?

Yes. A Skill's Allowed MCP tools list restricts the tool registry for the turn to only the listed tools. For tools from search connectors, each tool also has a permission of Always allow, Needs approval or Deny.

What is the difference between a skill and a workspace?

A Skill packages a method (prompt plus optional tool boundary). A Workspace packages a context (framing, default model, members and composed Skills). You usually compose Skills into workspaces.

How do I connect an external MCP server to Safeguard Search?

Open the connector dialog in Search. Use Browse Directory for listed servers and choose MCP server, OAuth 2.0 or personal access token. For anything else, use Custom MCP with a name, a full Streamable HTTP or SSE URL and an optional Authorization header.

Does the AI ask before calling a connector tool?

By default, yes. Every connector tool starts at Needs approval. You can change a tool or a whole category to Always allow or Deny.

Why does my workspace system prompt not seem to apply?

Check whether a Skill applied to the chat, or a setting on the conversation itself, is overriding it. Both take precedence over the workspace's system prompt and default model for that conversation.

Is a search connector the same as Safeguard's MCP server?

No. A search connector lets Griffin call tools on an outside MCP server. Safeguard's MCP server lets an outside assistant call Safeguard's tools. They are opposite directions with separate controls.

Fix the software your AI workflows investigate

Organizing AI-assisted work is about method, but the investigations themselves usually end at the same place: a vulnerable open source dependency or container image that needs to move to a safe version. The faster that fix lands, the less your triage workspace has to keep re-reporting the same vulnerability.

Fix it without waiting for upstream

Whatever the vulnerability, most of the risk sits in two gaps. The first is the time between a fix being published and that fix actually running in your production builds. The second, and worse, is the time before any fix exists at all, while a maintainer works out a patch and a release schedule. Waiting through both gaps is a habit, not a requirement. Safeguard gives teams three ways to close them.

1. Autonomous remediation: pull requests, not tickets

When a fixed version exists, Griffin, Safeguard's security model, opens a policy gated pull request that moves the affected dependency to the safe version, and your own CI runs your tests on it before anyone merges. You can run it on one repository or in bulk across every project that carries the vulnerable version. Remediation categories include dependencies, licenses, security posture and findings, and each one is set separately. Reachability analysis ranks the vulnerable code paths your application actually calls first, so the first pull requests you merge are the ones that matter most. Auto-merge is a separate switch and stays off by default, so your team decides what merges on its own and what still needs a human approval.

2. Gold Open Source: the safe version in seconds, free

gold.safeguard.sh is free and needs no login. Search any CVE or package to see affected and fixed versions, CVSS, EPSS, CISA KEV status and a recommended safe version, or use the "Am I affected?" checker. The same data is available through a free read API, README status badges and RSS threat feeds.

3. Gold Registry: start from components that are already fixed

Safeguard Gold rebuilds patched versions of open-source packages and images from source, signs each build, attaches an SBOM, SLSA provenance and VEX, and serves them from registry.safeguard.sh to your organization. Every Gold artifact is held continuously to zero exploitable critical or high CVEs and a completed license check, with signed build provenance (signed with Safeguard's key and verifiable with cosign, SLSA Build Level 3 as the design target) and a public evidence page you can hand to an auditor. An artifact loses its Gold badge if it regresses. Teams can point their own package directory at the registry so hardened equivalents resolve without changing how developers install anything.

When upstream has no fix yet, or will not ship one, Safeguard first works to get the fix accepted upstream. If the issue is severe or upstream will not move, Safeguard builds a fork, fixes the vulnerability there and publishes a hardened Gold build, so you are not stuck on a maintainer's release schedule. You can also request a hardened build of a specific package. Ask which packages it covers before you plan around it.

Your situationWhat to use
A fixed version existsAutonomous remediation opens the upgrade pull request; Gold Open Source confirms the safe version
You need to know today whether you are affectedGold Open Source lookup and "Am I affected?" checker, free and without login
You want vulnerable versions to stop shippingSafeguard policies and gates in CI
Upstream has no fix yet, or will not ship oneA custom Gold request; fork and patch
You want to start clean on new projectsPull hardened components and images from Gold Registry

One honest limit: when the vulnerable code lives in a vendor appliance, a hosted service or an operating system runtime you do not build yourself, the fix still has to come from that vendor or distribution. The steps above apply to the open source dependencies and container images you build and ship, which is where most teams carry the largest and least tracked share of their exposure.

See how Safeguard would help

The first demo connects Claude to the Safeguard MCP server with OAuth and shows choosing which tools Claude can use. It is the inward direction described above, and the same habit applies to Search Connectors: decide deliberately which tools an assistant may call before it calls them.

Click through it at your own pace, or open it full screen.

The second demo is a real Safeguard CLI session recorded on a Mac on 21 October 2026: install, ask for a fix, approve each step, then scan, on a small sample project with test data only. It shows the same habit as connector permissions, a person approving steps, in the terminal.

Click through it at your own pace, or open it full screen.

To try it on your own projects, download Safeguard or start with the Safeguard CLI docs.

Quick check

A Skill has an empty Allowed MCP tools list. Which tools can the model call on that turn?

No extra restriction applies, so the turn's tool registry is unrestricted. To bound the tools, list each one by its exact registered name.

A workspace sets a default model, and an analyst applies a Skill whose instructions conflict with the workspace. Which wins?

The Skill takes precedence for the conversation it is applied to. A setting on the conversation itself would also override the workspace.

Sources

Last reviewed: 28 October 2026.

Never miss an update

Weekly insights on software supply chain security, delivered to your inbox.

We email you a link to confirm, and only then add you to the list. Unsubscribe in one click. Privacy

Self-healing security runs on Safeguard.

Your first fix PR is minutes away.

No sales call required, even your agent can complete the purchase over MCP.