Native provider limits
17%
DETECTION COVERAGE
0% run-scoped · $6.62 exposure
OPEN BENCHMARK · V0.1.0
Twelve ways software runs away.
RunawayBench injects recursive agents, retry storms, queue cycles, fallback loops, and margin attacks into deterministic workloads. The LimitFuse rows execute the same policy engine used by the gateway.
SCOPED ENFORCEMENT
100%
All 12 scenarios detected and blocked at one run, with healthy control traffic unaffected.
Framework middleware
50%
DETECTION COVERAGE
0% run-scoped · $5.74 exposure
Gateway spend controls
17%
DETECTION COVERAGE
0% run-scoped · $6.62 exposure
LimitFuse shadow detection
100%
DETECTION COVERAGE
0% run-scoped · $8.64 exposure
LimitFuse scoped enforcement
100%
DETECTION COVERAGE
100% run-scoped · $2.06 exposure
SCENARIO MATRIX
What each control sees—and what it stops.
“Native,” “framework,” and “gateway” are published reference capability profiles, not claims that every product in a category behaves identically. Cost is modeled exposure in this synthetic run.
| FAILURE MODE | NATIVE PROVIDER LIMITS | FRAMEWORK MIDDLEWARE | GATEWAY SPEND CONTROLS | LIMITFUSE SHADOW DETECTION | LIMITFUSE SCOPED ENFORCEMENT |
|---|---|---|---|---|---|
01 · Recursive agent delegation agent · depth | missed· $0.50 | process· $0.10 | missed· $0.50 | shadow· $0.50 | run· $0.20 |
02 · Repeated malformed tool output agent · retry | missed· $0.72 | process· $0.16 | missed· $0.72 | shadow· $0.72 | run· $0.13 |
03 · Infinite webhook retry webhook · retry | missed· $0.42 | process· $0.07 | missed· $0.42 | shadow· $0.42 | run· $0.06 |
04 · Queue visibility-timeout storm queue · duplicate | missed· $0.54 | missed· $0.54 | missed· $0.54 | shadow· $0.54 | run· $0.11 |
05 · Function writing to its own trigger function · structural-loop | missed· $0.55 | missed· $0.55 | missed· $0.55 | shadow· $0.55 | run· $0.13 |
06 · Duplicate event delivery cross-layer · duplicate | missed· $0.40 | missed· $0.40 | missed· $0.40 | shadow· $0.40 | run· $0.10 |
07 · Long-context growth model · context-growth | provider-account· $0.20 | missed· $1.98 | project· $0.20 | shadow· $1.98 | run· $0.90 |
08 · Unbounded parallel research agent · fanout | missed· $0.55 | process· $0.11 | missed· $0.55 | shadow· $0.55 | run· $0.09 |
09 · Model fallback cycle model · fallback | missed· $0.84 | process· $0.14 | missed· $0.84 | shadow· $0.84 | run· $0.11 |
10 · Malicious user cost amplification cross-layer · unit-economics | provider-account· $0.96 | missed· $1.20 | project· $0.96 | shadow· $1.20 | run· $0.10 |
11 · Tool timeout with partial success agent · partial-success | missed· $0.58 | process· $0.13 | missed· $0.58 | shadow· $0.58 | run· $0.10 |
12 · Queue consumer poison message queue · poison-message | missed· $0.36 | missed· $0.36 | missed· $0.36 | shadow· $0.36 | run· $0.05 |
REPRODUCIBILITY
Read the method before the score.
Twelve deterministic failure scenarios from the LimitFuse product blueprint; no external provider calls.
LimitFuse profiles execute the production policy evaluator and integer-microdollar wallet implementation.
Native, framework, and gateway rows are transparent reference profiles, not claims about every vendor implementation.
Cost before decision uses 80% realized cost for admitted LimitFuse operations; uncontrolled/reference exposure uses declared worst-case operation cost.
Scoped containment requires a blocking run-level decision that leaves the healthy control run unaffected.