v3.1 (1/N): extract ssrf-guard-llm + add ssrf-guard-langchain4j#6
Merged
Conversation
Splits the LLM-agent SSRF defense across two concerns:
1. ssrf-guard-llm (new module) — framework-agnostic core that holds the
JSON walking + URL extraction + policy validation logic. Exposes
ToolInputGuard (interface) and JsonToolInputGuard (default impl).
2. Adapter modules become ~30-line thin wrappers that translate between
each LLM framework's own ToolExecutor/ToolCallback interface and the
neutral ToolInputGuard contract.
Refactored:
- ssrf-guard-springai: SsrfGuardedToolCallback delegates to
JsonToolInputGuard via the new -llm api. Public API unchanged
(constructors and methods keep the same shape) — v3.0.x consumers
pick up -llm transitively and see no behavioural change. All 10
existing springai tests pass.
Added:
- ssrf-guard-langchain4j: SsrfGuardedToolExecutor wraps a LangChain4j
ToolExecutor, inspects ToolExecutionRequest.arguments() through
the same JsonToolInputGuard. Auto-wrap via BeanPostProcessor for
Spring Boot consumers; SsrfGuardedToolExecutors.wrap(...) helpers
for non-Spring / programmatic use. Targets LangChain4j 1.15.0 (the
first GA line).
Why split this way
Previously the JSON walk + URL detection + error formatting lived
inside SsrfGuardedToolCallback (~200 lines). Adding LangChain4j the
naive way meant copying those 200 lines verbatim into
SsrfGuardedToolExecutor and fixing bugs in two places forever.
With the extraction, a fix to URL detection or error formatting lands
in one place (ssrf-guard-llm). New frameworks (MCP, custom tool
dispatchers, future Mistral/Cohere SDKs) get ~30 lines of adapter
work, not 200.
User experience is unchanged:
- Spring AI users: install ssrf-guard-springai, declare regular
@bean ToolCallback, autoconfig BeanPostProcessor wraps it.
- LangChain4j users: install ssrf-guard-langchain4j, declare regular
@bean ToolExecutor, autoconfig BeanPostProcessor wraps it.
- No user code touches ToolInputGuard unless they want to roll their
own dispatcher.
Build state
All 199 tests across 11 modules pass (was 182 + 17 new). Local full
build green. Maven publish coordinates declared but not yet released
— that's Phase 6 (CHANGELOG + tag).
Remaining v3.1 phases (separate PRs):
- Phase 4: WebClient reactor-netty DNS-time guard
- Phase 5: GraalVM Native Image hints
- Phase 6: CHANGELOG + tag v3.1.0
3 tasks
jlc488
added a commit
that referenced
this pull request
May 23, 2026
Cleans up the v3.1 baseline before tagging:
Docs (en/ko, symmetric):
- index.md / .ko.md — module matrix gets -llm and
-langchain4j rows; webclient row
notes the new URL+DNS coverage.
- README.md / .ko.md — same matrix in the repo's front-door.
- installation.md / .ko.md — two new install tabs (LangChain4j
tool execution, Custom tool
dispatcher via ssrf-guard-llm).
- guides/configuration.md/.ko.md — new "LLM-adapter properties"
section documenting
`ssrf.guard.springai.wrap-tool-callbacks`
and the new
`ssrf.guard.langchain4j.wrap-tool-executors`.
- guides/security-model.md/.ko.md — the "what this doesn't protect"
bullet was stale (mentioned
RestClient only); now covers
every module v3.1+ ships. Added
a dedicated bullet about the
WebClient URL-time → URL+DNS
upgrade.
Autoconfig integration tests:
- SsrfGuardLangchain4jAutoConfigurationTest — boots a Spring context
with the autoconfig, declares a consumer @bean ToolExecutor, then
asserts:
* the BeanPostProcessor wrapped it (instanceof check)
* end-to-end block on an AWS metadata URL (LLM sees the
structured JSON error)
* end-to-end allow on a whitelisted URL (delegate's
PRETEND-FETCHED echo)
* `wrap-tool-executors=false` correctly leaves the bean
unwrapped — the off switch for manual-wrap workflows.
- SsrfGuardWebClientAutoConfigurationTest extended to verify the
new v3.1 beans (ssrfReactorClientHttpConnector,
ssrfReactorAddressResolverGroup) are registered. Previously the
test only checked filter / customizer / builder; the new beans
silently going missing in a future refactor would have flown
under the radar.
Tests now total 210 across 11 modules (up from 199 in PR #6 and
204 after PR #7).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
First of several v3.1 PRs. Extracts the LLM-agent SSRF defense's reusable core into a new framework-agnostic module and adds a second adapter for LangChain4j.
What changed
New:
ssrf-guard-llmFramework-agnostic core. Holds the JSON walking + URL extraction + policy validation that every LLM tool adapter shares. Exposes:
ToolInputGuard(interface) — the neutral contract:String checkOrFormatError(String toolInput)JsonToolInputGuard(default impl) — Jackson-based JSON walker, validates everyhttp(s)://string it finds at any depth againstUrlPolicy. Optionally throws on violation (CI mode) or returns a structured JSON error string (LLM-recovery mode).New:
ssrf-guard-langchain4jAdapter wrapping LangChain4j's
ToolExecutor.SsrfGuardedToolExecutorinspectsToolExecutionRequest.arguments()throughJsonToolInputGuard.BeanPostProcessor(mirrorsssrf-guard-springai's pattern).SsrfGuardedToolExecutors.wrap(map)/wrap(collection)/wrapOne(executor)helpers for non-Spring / programmatic use (theMap<ToolSpecification, ToolExecutor>shape AiServices expects).Targets LangChain4j 1.15.0 (current GA line).
Refactored:
ssrf-guard-springaiSsrfGuardedToolCallbackshrunk to ~30 lines — delegates toJsonToolInputGuardfrom-llm. Public API unchanged — same constructors, same methods, same behaviour. v3.0.x consumers see no API change; they just pick upssrf-guard-llmtransitively.Why this shape
User asked the right question: "If two adapters have the same implementation and only the framework interface differs, isn't a common core + thin adapters the better design than copying ~200 lines twice?"
Yes. With the extraction:
SsrfGuardedToolCallback(~200 lines)JsonToolInputGuardin-llm(single source)The new
ToolInputGuardinterface also gives advanced users (custom dispatchers, MCP servers, internal tool routers) a clean hook — they DIJsonToolInputGuardand callcheckOrFormatError(input)themselves. Optional, not required.Non-breaking for v3.0.x consumers
ssrf-guard-springai's public types (SsrfGuardedToolCallback,SsrfGuardedToolCallbacks,SsrfGuardSpringAiAutoConfiguration) keep the same signatures. The only observable difference: an extra transitive dep onssrf-guard-llm(which pulls Jackson, same as before).Test plan
:ssrf-guard-llm:test— 9 tests covering JSON walk, nested URLs, arrays, non-JSON input, throw-on-violation:ssrf-guard-springai:test— 10 existing tests still pass after refactor:ssrf-guard-langchain4j:test— 7 tests mirroring springai's threat model on the LangChain4j abstractionWhat's NOT in this PR (follow-ups)