| Spoofing | A look-alike repository or package distributes a tampered corpus | Every blob is named by its sha256 and the manifest records sha256 and byte_length per variant; release archives carry checksums (corpus.json checksums); xdrcheck-corpus verify --checks hashes re-checks every blob | A consumer who skips verification |
| Tampering | A contribution changes bytes or expectations so that a real bug passes or a correct reader fails | Released bytes are immutable (maintenance.md §9); every pull request re-hashes every blob, sidecar and vendored object (make repo-checks, make verify-corpus); golden entries must equal the named frame of the file at the pinned stellar-core tag (ag1); local-core entries must regenerate (make verify-generate); expectations must agree with the reference extractor (extractor check); a reviewer who did not write the entry answers the review checklist (authoring-entries.md §6) | Expectations are the weak point. Bytes are protected by hashes; a wrong hand-curated expectation on a history entry is checked only by the reference extractor and a reviewer. Partial mitigations: pinpoints written independently of the extractor, and the fact that a wrong expectation shows up as the same divergence in every SDK column at once |
| Tampering | Supply chain: a pinned SDK, image, package or download is replaced upstream | Every version in pins.toml, images by digest, downloads by sha256, lock files with integrity hashes; xdrcheck pins check fails on any version named elsewhere; npm installs run with --ignore-scripts; third-party CI actions pinned by commit; make fetch verifies hashes before anything is used | A compromised upstream release that is then pinned on purpose; every pin move carries a differential report for review |
| Repudiation | Dispute over what a run observed | Reports (xcc.report/v1) record tool versions, pins, entry hashes and the observed values next to the expected ones; CI uploads them as artefacts | — |
| Information disclosure | Secrets in the repository or CI | No secret is needed by any job; workflows run with permissions: contents: read and persist-credentials: false; scripts/py/check_public.py (in make repo-checks) rejects machine paths, private names and internal references | — |
| Denial of service | The corpus contains, on purpose, inputs built to stress readers (deep nesting, large length prefixes, very large ledgers). Run in another project’s CI they can exhaust memory, stack or time | The manifest flags negative entries with a reason class; release archives keep stress entries in a separate directory, out of the default run (release.md); the kit’s own runs decode with explicit limits and run every driver and adapter under memory and time limits (ci-isolation.md, xdrcheck conform run --memory-max, xdrcheck replay --memory-max); stress entries run in a fresh driver process each | A consumer who opts in to the stress entries and decodes them without limits — which is the failure the corpus exists to expose |
| Elevation of privilege | Code execution in CI through a contribution | Scenario files and recipes are declarative data validated against schemas. Three things that run are code: the test contracts’ wasm (executed by stellar-core on a local chain), the adapters and the drivers; third-party SDKs and the pipelines under test are code too. Controls: contracts are built reproducibly from reviewed source and the build fails if the wasm hash differs (contracts/build.sh); drivers and adapters change only through reviewed pull requests; pull requests from forks get no secrets or write token; every run step executes offline (OFFLINE=1) under resource limits | A reviewed-in malicious change; a compromised pinned package executing inside the sandbox, where it can reach only the checkout |
| Misuse | The corpus publishes transaction envelopes signed under the public network passphrase with keys anyone can derive; they could be mistaken for real transactions or replayed | They come from a private standalone network and are labelled as local-core in the manifest. They cannot be replayed on the public network because of sequence numbers: an account created on the local chain at ledger L has a sequence number near L << 32, while every account on the public network today was created tens of millions of ledgers later and its sequence number only rises, so the envelopes fail with a bad sequence number. The README warns never to fund a label-derived account | (1) The root account’s key derives from the network passphrase, so on the real networks it is the genesis account; scenarios use root only for initial funding. (2) After a testnet reset sequence numbers start again, so testnet-passphrase envelopes may become valid on the new testnet for a while; only test funds are involved. (3) Anyone who funds a label-derived account on a real network can lose those funds to anyone else |
| Licence | Vendored third-party bytes and code are redistributed | Lake objects and SEP vectors are public data; stellar-core goldens and settings are Apache-2.0; each source is listed in NOTICE and its provenance recorded per entry | — |