ADR-0040: Reject ReDoS-prone inputSchema patterns with a load-time lint
Context and Problem Statement
An agent's optional inputSchema is a JSON Schema (Draft 2020-12) that Homer validates the request body against. InputSchemaValidator delegates that validation to the networknt json-schema-validator, which compiles each schema pattern to java.util.regex and matches it against caller-supplied body values. A pattern that backtracks catastrophically (e.g. ^(a+)+$) combined with a crafted small value pins a request thread's CPU indefinitely — a ReDoS denial-of-service reachable by any caller allowed to POST a body.
ParamValidator already bounds the identical "author pattern + caller value" vector for inputs params via BoundedRegex, which runs the match on a worker thread reading an InterruptibleCharSequence and interrupts it on timeout. InputSchemaValidator has no such bound.
Decision Drivers
- The vulnerable regex is author-supplied and evaluated against attacker-influenceable input, so the guard must be structural or fail-closed.
- The fix must be contained and backward-compatible — it must not change how legitimate schemas validate bodies.
- Homer does not own the match: networknt compiles the pattern internally and feeds it a plain
String, so theBoundedRegexinterruption trick cannot be injected here.
Considered Options
- Load-time lint — reject schemas whose
patternexhibits classic evil-regex shapes when the agent is loaded. - Runtime wall-clock timeout around networknt's
validatecall. - Swap the regex engine for a linear-time one (RE2J).
Decision Outcome
Chosen option: load-time lint. InputSchemaValidator.precompile (already invoked at agent load by AgentsService, and once per distinct schema) walks the schema for every pattern string and patternProperties key and rejects any that RedosPatternDetector flags — nested unbounded quantifiers ((X+)+, (X*)*, (X+)*, (X*)+, ((X+))+) and quantified alternations with overlapping branches ((a|a)*, (a|ab)*, (a|)*), including alternations that are non-capturing ((?:a|ab)*) or wrapped in nested groups (((a|ab))*) — the detector normalises (?…) group prefixes and carries an overlap up through nesting. A flagged pattern fails agent load with a message naming the pattern, exactly like a malformed schema.
The detector is a documented heuristic: conservative, neither sound nor complete. It can over-reject an unusual-but-safe pattern and can miss an exotic vulnerable one; an author who trips it restructures the pattern.
Consequences
- Good: the vulnerable pattern is rejected before any match runs, so no crafted body ever reaches an unbounded matcher. No per-request cost.
- Good: no new dependency, no engine swap, no behavioral change for schemas that already validate bodies today.
- Bad: heuristic — false positives (a safe pattern rejected) and false negatives (an exotic evil regex slipping through) are both possible.
- Bad: covers
inputSchemapatterns only;patternPropertieskeys are covered, butformat/$ref-driven remote regexes are out of scope.
Pros and Cons of the Options
Runtime wall-clock timeout
- Good: catches any vulnerable pattern regardless of shape.
- Bad: a naive
Future.get(timeout)does not stop the CPU burn — ajava.util.regex.Matcherin a backtracking loop ignores interruption unless it re-reads an interruptibleCharSequence(howBoundedRegexworks). networknt owns the match and feeds it a plainString, so we cannot inject that; the worker thread keeps burning after the timeout returns. - Bad: per-request cost and a leaked runaway thread.
Swap the regex engine (RE2J)
- Good: linear-time matching eliminates the vector entirely.
- Bad: networknt is coupled to
java.util.regex; swapping means replacing or forking the validator. RE2J also rejects somejava.util.regexconstructs (backreferences, lookaround), silently changing validation semantics for existing schemas — not backward-compatible.
More Information
Related: BoundedRegex, ParamValidator (the inputs equivalent, ADR-0031).