Supply-chain, crypto, events, and data-boundary batch¶
Source: GitHub Security Advisories updated 2026-05-13.
Package compromise and data-format advisories are durable because they live between trust domains: package registries, event streams, blockchain certificates, HTML helpers, and broker/client semantics.
Advisories covered¶
- PyTorch Lightning PyPI package compromise — GHSA-w37p-236h-pfx3: published package versions were compromised, requiring dependency, token, and build-host triage.
- sse-channel event-field injection — GHSA-84hm-wfh8-c5pg: unsanitized event fields could inject Server-Sent Event semantics.
- ip-address Address6 HTML XSS — GHSA-v2v4-37r5-5v8g: HTML-emitting helpers could render address data as scriptable output.
- bsv-sdk ARC broadcaster success confusion — GHSA-9hfr-gw99-8rhx: INVALID, MALFORMED, or ORPHAN responses were treated as successful broadcasts.
- bsv-sdk and bsv-wallet unverified certifier signatures — GHSA-hc36-c89j-5f4j: certificate acquisition paths persisted unverified certifier signatures.
- RoadRunner HTTP request/response smuggling — GHSA-g9pc-8g42-g6vq: a vulnerable dependency created request/response boundary ambiguity.
Operator triage¶
- Patch affected packages and hosted services first where the vulnerable component is internet-facing, tenant-facing, or reachable by untrusted project data.
- Inventory transitive exposure; many of these bugs live in helpers, plugins, middleware, scanner images, or framework defaults rather than application code.
- Search logs for boundary probes: encoded paths, unusual headers, oversized bodies, duplicate auth attempts, symlinked project files, private-network URLs, and stored HTML/script payloads.
- Add regression tests at the trust boundary, not only at the direct vulnerable function. Exercise canonicalized paths, redirects, alternate address syntax, concurrent auth, and malformed protocol inputs.
Durable controls¶
- Canonicalize once, authorize after canonicalization, and execute/use only the canonicalized object.
- Give every parser, helper, cache, upload, range handler, and HTTP client explicit byte, item, time, and recursion budgets.
- Treat user-controlled templates, package metadata, project files, identity headers, event fields, and backup archives as untrusted code-adjacent inputs.
- Prefer positive allowlists tied to resolved identities/resources over deny-lists tied to raw input strings.