Security fixes are applied to the default branch (main) and released as semver tags (vMAJOR.MINOR.PATCH). Treat the latest GitHub release as the current supported line.

• Latest release: full support (including security fixes).
• Older tags: we do not guarantee long-term support for every past version; upgrade to the latest release when possible.

Maintainers document breaking changes in release notes. See RELEASING.md for how versions are cut.

Please report security vulnerabilities by opening a GitHub Security Advisory. Do not open a public issue for security vulnerabilities.

We will acknowledge your report within 48 hours and will send a more detailed response within 7 days. Please do not publicly disclose the vulnerability until we have released a fix.

We appreciate responsible disclosure and will acknowledge security researchers who help us improve the security of this project (with their permission). We do not currently offer a bug bounty or monetary rewards for vulnerability reports.

• Security issues in this SDK (agent, tools, conversation, LLM clients)
• Sensitive data exposure (API keys, approval payloads)

When deploying agents built with this SDK, review the following application-level risks; the SDK provides the hooks below, but callers own validation and access control.

Agent tools execute based on LLM decisions. Validate and sanitize all external data before injecting into agent context. Use WithHooks(name, AgentHooks) with a BeforeLLM hook to inspect and filter inputs.

When prompt caching is enabled (Anthropic), prefix KV state is stored server-side at the provider. Leave caching disabled (default) or disable explicitly with llm.WithPromptCaching(false).

MCP stdio transport spawns local processes. Only register MCP servers from trusted sources. Validate MCP server binaries before registering via WithMCPConfig().

Never pass API keys through agent prompts or tool inputs. Use environment variables or secret managers. Keys may be held in process memory via llm.WithAPIKey; the SDK does not persist them to disk or put them in prompts.

Implement interfaces.ToolAuthorizer (Authorize) to enforce access control before tool execution. For human-sensitive operations, implement interfaces.ToolApproval (ApprovalRequired) and/or use WithToolApprovalPolicy with WithApprovalHandler.

Key dependencies — monitor via Dependabot (.github/dependabot.yml, weekly PRs for / and /cli) and govulncheck (task govuln, Security workflow, and Release hard-fail before GoReleaser):

• go.temporal.io/sdk
• github.com/restatedev/sdk-go
• github.com/modelcontextprotocol/go-sdk
• github.com/a2aproject/a2a-go/v2

• Temporal server or Temporal Cloud
• Restate server or Restate Cloud
• Third-party LLM providers (OpenAI, Anthropic, Google)
• General usage questions