CVE Database

SGLang through 0.5.19 in prefill/decode disaggregation mode contains an unauthenticated PUT /route endpoint on the prefill bootstrap service that allows attackers to poison the KV transfer routing table. Attackers can supply arbitrary rank_ip and rank_port values to redirect decode workers to attacker-controlled endpoints, causing denial of service or disclosure of KV transfer metadata including session identifiers and tensor-parallel topology parameters.

InternLM LMDeploy through 0.17.0 contains a reachable assertion vulnerability in the DistServe decode migration loop that allows unauthenticated attackers to terminate the inference engine. Attackers can submit a migration_request with an empty remote_block_ids list to trigger an AssertionError that crashes the engine loop and causes subsequent inference requests to fail.

HUBzero CMS through 2.2.32 contains a path traversal vulnerability in project file upload handlers that allows authenticated project members to write arbitrary files outside the project repository. Attackers can supply traversal sequences in upload parameters to write files to attacker-chosen paths with web server privileges, potentially enabling code execution.

vm2 versions before 3.11.2 fail to properly restrict access to the VM2_INTERNAL_STATE_DO_NOT_USE_OR_PROGRAM_WILL_FAIL global variable. Attackers can access this internal state object through globalThis to retrieve sensitive sandbox internals.

vm2 before 3.11.6 fails to enforce bufferAllocLimit on ArrayBuffer, SharedArrayBuffer, and TypedArray constructors, allowing attackers to allocate arbitrary host memory. Attackers can bypass the buffer allocation cap by using these V8 intrinsics to exhaust host process memory and trigger out-of-memory conditions.

vm2 before 3.11.8 does not fully enforce the allowAsync: false option in VM and NodeVM. While localPromise.prototype.then is replaced with a handler that throws 'Async not available', the sandbox's Promise static methods (Promise.resolve, Promise.all, Promise.race, Promise.any, and Promise.allSettled) still assimilate attacker-supplied thenables: native promise resolution performs PromiseResolveThenableJob and invokes the sandboxed code's then method in a microtask without passing through the patched then, so the async restriction is never applied. As a result, sandboxed script can schedule work that runs after VM.run() or NodeVM.run() has returned and outside the configured timeout, continuing to execute after the host believes execution is complete.

vm2 through 3.11.6 contains a builtin-module denylist bypass in NodeVM. When the embedder uses the builtin wildcard together with negative entries (e.g. require: { builtin: ['*', '-fs', '-child_process'] }), negative entries are matched by exact module name in lib/builtin.js, so -fs removes only the builtin named fs and does not remove builtin subpaths such as fs/promises. Sandboxed code can therefore call require('fs/promises') or require('node:fs/promises') and reach the promise-based filesystem API despite fs being denied; node: prefix handling is likewise inconsistent (a -node:fs/promises entry does not block require('fs/promises')). Host file creation and writing were confirmed via fsp.writeFile(), and other fs/promises operations (cp, mkdir, rename, rm, rmdir, truncate, read operations, etc.) are also reachable. This issue is fixed in vm2 3.11.7.

vm2 through 3.11.6 does not normalize `node:`-prefixed builtin specifiers when evaluating user-supplied negative (deny) entries in a NodeVM wildcard require policy. Although NodeVM strips the `node:` prefix during require() resolution, negative wildcard entries are matched by exact string comparison against the canonical builtin names, so a policy such as `new NodeVM({ require: { builtin: ['*', '-node:child_process'] } })` fails to deny the canonical `child_process` module. Sandboxed code can therefore obtain the host `child_process` builtin via `require('child_process')` or `require('node:child_process')`, gaining references to process-spawning APIs such as execSync and spawn, which is equivalent to host command-execution capability for untrusted sandbox code. Fixed in vm2 3.11.7.

vm2 is a sandbox library for running untrusted JavaScript in Node.js. In versions >= 3.10.0 and <= 3.11.7, Promises returned from the host realm into the sandbox are not marked as handled at the bridge boundary; only Promises created inside the sandbox are wrapped with a rejection-swallowing handler (lib/setup-sandbox.js), and the bridge only installs host-side rejection sanitizers when sandbox code calls .then/.catch/.finally. As a result, code running in the sandbox can invoke a host function that returns a rejected Promise (for example events.once() exposed via the NodeVM events builtin, or any embedder-provided Promise-returning API) and simply ignore the return value, leaving the host Promise unhandled so that Node.js's default unhandled-rejection behavior terminates the host process. This is an incomplete fix of GHSA-hw58-p9xv-2mjh. The issue is fixed in version 3.11.8.

vm2 versions from 3.11.0 before 3.11.8 fail to protect host TypedArray and ArrayBuffer prototypes from sandbox mutation. Attackers can use prototype-walking primitives to reach and modify host Uint8Array.prototype, %TypedArray%.prototype, and ArrayBuffer.prototype, causing host-created typed arrays to observe attacker-controlled properties after VM.run() returns.

vm2 versions from 3.9.6 before 3.11.7 fail to properly restrict access to accessor properties on frozen objects, allowing sandboxed scripts to bypass vm.freeze() and vm.readonly() protections. Attackers can use Object.getOwnPropertyDescriptor() or __lookupSetter__() to extract and invoke host object setters directly, mutating properties the embedder explicitly marked read-only.

vm2 versions >= 3.9.6 and <= 3.11.6 are affected by a NodeVM builtin allowlist bypass that permits a sandbox escape on Node.js 24 and newer when the embedder explicitly allows the node:test builtin (e.g. require: { builtin: ['node:test'] }). On Node.js 24+, module.builtinModules exposes the scheme-only key node:test, which is not covered by vm2's family-based DANGEROUS_BUILTINS protection, so it is stored in the generic host-passthrough loader. Because requireImpl() in lib/setup-node-sandbox.js strips a single 'node:' prefix before the builtin lookup, sandbox code calling require('node:node:test') resolves to the stored node:test key and receives a readonly proxy to the host module. Calls to node:test.run() are forwarded to the host implementation, which spawns a separate Node process for process-isolated test execution and passes through attacker-controlled execArgv values; supplying --eval=<JavaScript> therefore executes arbitrary JavaScript in an unrestricted host Node process outside the NodeVM sandbox. Fixed in vm2 3.11.7.

vm2 versions 3.10.2 through 3.11.6 contain a sandbox escape vulnerability on Node.js 26 where Promise.prototype.finally() bypasses vm2's wrapper protections due to a stale PromiseThenLookupChain protector in V8 14.6. Attackers can exploit this by creating an async function that returns a Promise with an attacker-controlled constructor Symbol.species, allowing them to reach the host Function constructor and process object for arbitrary code execution.

vm2 before 3.11.7 (affected versions <= 3.11.6) does not enforce the VM({ timeout }) option on code executed outside the synchronous VM#run() call. The timeout only wraps the single call to _runScript() via doWithTimeout() in lib/vm.js, and FinalizationRegistry and WeakRef are exposed to sandboxed code unmodified (they are not among the hardened globals in lib/setup-sandbox.js). Sandboxed code can register a FinalizationRegistry cleanup callback against an object and then drop the only strong reference to it; vm.run() returns within the configured timeout, but when the V8 garbage collector later reclaims the object it invokes the sandboxed cleanup callback outside any vm2 timeout accounting. A busy loop in that callback blocks the host event loop for an unbounded period, resulting in denial of service. The time of invocation depends on the garbage collector (e.g. under memory pressure or with --expose-gc).

vm2 versions 3.11.3 through 3.11.6 expose Node.js's host node:sqlite module to code running in NodeVM when that builtin is permitted, either explicitly or through builtin: ['*']. The module is wrapped with vm.readonly(), which prevents property assignment but leaves host-authority callables reachable; in addition, the resolver treats any request starting with 'node:' as a core-module request and the runtime strips only one 'node:' prefix, so a sandbox request for 'node:node:sqlite' resolves to the configured node:sqlite entry. Sandboxed code can therefore create an in-memory DatabaseSync with extension loading enabled and call DatabaseSync.loadExtension() on a native library bundled in the untrusted plugin package (path derived from __dirname). SQLite loads the library into the Node.js host process and invokes its native entry point, giving the sandboxed plugin arbitrary native code execution outside the sandbox with the host process's privileges. The issue is fixed in vm2 3.11.7.

vm2 3.11.6 is vulnerable to a sandbox escape leading to remote code execution in the host Node.js process. The fix for GHSA-m283-3h24-438v is incomplete: the bridge gate at lib/bridge.js:1624 identity-checks only the direct call target when deciding whether to rebuild/sanitise a rejected host Promise value. Registering the rejection handler through Function.prototype.call or .apply indirection (e.g., p.then.call(p, undefined, cb)) makes the intercepted target host Function.prototype.call, so the sanitiser never runs and the raw host error reaches sandbox code with its own properties intact. If an embedder exposes a host-realm Promise to the sandbox (an async host function bridged via the sandbox option, or a NodeVM external module's async method) and that Promise rejects with an Error carrying a non-primitive own property referencing a host object (for example err.detail = process), untrusted code in the sandbox obtains a fully functional proxy to that host object and can execute arbitrary commands with the privileges of the host process (e.g., e.detail.mainModule.require('child_process').execSync(...)). The direct p.then(undefined, cb), bind, and Reflect.apply forms are correctly sanitised. Fixed in vm2 3.11.7.

vm2 is a sandbox for running untrusted Node.js code. In versions <= 3.11.7, NodeVM exposes the host `util` module to the sandbox as an unfiltered shallow copy (`Object.assign({}, util)` in `defaultBuiltinLoaderUtil`), and the deprecated `sys` builtin (an alias of host `util`) is exposed through the generic builtin loader. On Node.js >= 22.9 this hands sandboxed code `util.getCallSites()`, a programmatic stack-introspection API that returns the host process's full call stack, including absolute file paths, function names, and line numbers for vm2 bridge internals and the embedding application's entrypoint. This bypasses the host-frame redaction introduced for GHSA-v27g-jcqj-v8rw, which only applies to the `Error.prepareStackTrace` formatting channel. The issue is fixed in vm2 3.11.8.

google-protobuf contains an unbounded recursion when parsing unknown protobuf group fields. An attacker can send a small crafted payload of deeply nested START_GROUP wire bytes to any Node.js service that calls the generated deserializeBinary() API, causing a RangeError: Maximum call stack size exceeded and crashing the process. No authentication or prior knowledge of the schema is required.

Insufficient Session Expiration vulnerability in team-alembic AshAuthentication and AshAuthentication Phoenix allows a revoked session to remain fully authenticated. A resource configured with session_identifier :jti and require_token_presence_for_authentication? disabled stores its session value as <jti>:<subject>. The jti is there so that signing out can revoke that one session. Neither reader consults it: AshAuthentication.Plug.Helpers.authenticate_resource_from_session/4 and AshAuthentication.Phoenix.LiveSession.on_mount/4 both split the value with split_identifier/2, discard the jti and pass the bare subject to AshAuthentication.subject_to_user/3, which reloads the record. The token-presence branch of each function does check its token, calling AshAuthentication.TokenResource.Actions.get_token/3 with the jti and the purpose user. Because the revocation record is never read, neither its revoked state nor its expiry constrains the session, so a session captured before sign-out keeps working. This issue affects ash_authentication: from 4.9.1 before 4.15.0 and from 5.0.0-rc.0 before 5.0.0-rc.14; ash_authentication_phoenix: from 2.10.0 before 2.17.4 and from 3.0.0-rc.0 before 3.0.0-rc.11.

Improper Output Neutralization for Logs vulnerability in team-alembic AshAuthentication allows an unauthenticated attacker to forge application log entries by submitting a password reset identity containing newlines or control characters. AshAuthentication.Strategy.Password.RequestPasswordReset.run/3 interpolates the identity argument, the email or username taken straight from the reset request, into its Logger.warning/1 heredocs without escaping, truncating or type-restricting it. The resource logged beside it is passed through inspect/1, which would have neutralized the value. A newline in the identity therefore ends the log record, and everything after it is written as a line of its own, so an attacker chooses the severity tag and the content of entries that appear to have come from the application. This issue affects ash_authentication: from 4.2.0 before 4.15.0 and from 5.0.0-rc.0 before 5.0.0-rc.14.

Authentication Bypass by Primary Weakness vulnerability in team-alembic AshAuthentication allows an unconfirmed user to obtain a session, defeating a mandatory email confirmation requirement. AshAuthentication.Strategy.Password.Actions.check_user/2 decides whether the attribute named by require_confirmed_with is set using a bare is_nil(Map.get(user, value)). When that attribute is not selected on the loaded record Map.get/2 returns %Ash.NotLoaded{}, and when a field policy denies it for the current actor it returns %Ash.ForbiddenField{}. Neither is nil, so the rejection branch is skipped and sign-i require_confirmed_with is enforced in two places, and neither holds in every configuration. sign_in_with_token and register are checked only inside AshAuthentication.Strategy.Password.Actions, not on the action itself, so any caller that invokes the action directly skips the check. An API layer such as AshGraphql or AshJsonApi invokes the action directly, so this applies to the default configuration. Where a check does run it compares the confirmation attribute against nil. That attribute holds %Ash.NotLoaded{} or %Ash.ForbiddenField{} when it sets select_by_default?: false, when an API layer narrows the read's select, or when a field policy hides it from the sign-in actor. Neither struct is nil, so those configurations read every user as confirmed. This issue affects ash_authentication: from 4.3.8 before 4.15.0 and from 5.0.0-rc.0 before 5.0.0-rc.14.

Time-of-check Time-of-use (TOCTOU) Race Condition vulnerability in team-alembic AshAuthentication allows an attacker holding a leaked magic link to replay its single-use token and authenticate as the target subject. A magic link configured with single_use_token?, which is the default, is meant to be redeemable exactly once, but nothing serialises the token's validity check against its consumption, so concurrent redemptions of one token all succeed and each yields a full user token. Sign-in verifies the JWT with Jwt.verify/4 and revokes it only afterwards: AshAuthentication.Strategy.MagicLink.SignInPreparation revokes in a Query.after_action callback, and AshAuthentication.Strategy.MagicLink.SignInChange in an after_transaction hook that runs once the sign-in has already committed. AshAuthentication.TokenResource.Actions.revoke/3 writes the revocation as an upsert, so a concurrent duplicate revocation silently succeeds instead of conflicting and no request ever loses the race. This issue affects ash_authentication: from 3.9.0 before 4.15.0 and from 5.0.0-rc.0 before 5.0.0-rc.14.

Inefficient Algorithmic Complexity vulnerability in team-alembic AshAuthentication allows an unauthenticated attacker to exhaust CPU and memory via an oversized base62 segment in a submitted API key. AshAuthentication.Base.decode62/1 in lib/ash_authentication/base.ex splits its argument into one binary per character and folds it with charval62/2, which recomputes Integer.pow(62, index) at every position instead of accumulating by Horner's method, so cost grows roughly cubically in the input length. bindecode62/1 in the same module is quadratic through Integer.undigits/2 and Integer.digits/2. Neither function caps byte_size/1, and AshAuthentication.Strategy.ApiKey.SignInPreparation passes the underscore-separated segments of the submitted key straight into both, before any key lookup and without prior authentication. The surrounding rescue clauses catch exceptions, not CPU or memory exhaustion. This issue affects ash_authentication: from 4.8.0 before 4.15.0 and from 5.0.0-rc.0 before 5.0.0-rc.14.

Use of a One-Way Hash with a Predictable Salt vulnerability in team-alembic AshAuthentication allows readers of the audit store to recover the client IP addresses that the audit log add-on's :hash privacy mode is meant to pseudonymise. AshAuthentication.AddOn.AuditLog.IpPrivacy.hash_ip/1 computes a single unkeyed :crypto.hash(:sha256, salt <> ip) and truncates the result to 16 hexadecimal characters. The salt is read from the :audit_log_ip_salt or :secret application config keys, and falls back to the constant "default-salt-change-in-production" published in the library source when neither is set, with nothing warning that the default is in use. The IPv4 space is only 2^32 values and SHA-256 is fast, so the whole hash table is precomputable and every stored value maps back to its source address. Truncating to 16 characters does not help, and even a configured salt leaves the hash cheap enough to enumerate once it leaks. This issue affects ash_authentication: from 4.12.0 before 4.15.0 and from 5.0.0-rc.0 before 5.0.0-rc.14.

Insertion of Sensitive Information into Log File vulnerability in team-alembic AshAuthentication allows disclosure of user password digests to readers of the audit store. The audit_log add-on builds each entry's extra_data in AshAuthentication.AddOn.AuditLog.Auditor.build_extra_data/4, which takes :actor from the action callback context verbatim. Any audited action invoked with actor: set to a user record therefore deposits that record, including its hashed_password attribute, into the audit entry. The same module already collapses the audited identity to an opaque string via AshAuthentication.user_to_subject/1 and filters params against the strategy's configured allow-list, so the actor is the only stored value that reaches the audit store unfiltered. Marking the attribute sensitive?: true does not help, because that redacts inspect/1 output rather than JSON encoding or raw-term storage. There is no attacker-controlled trigger and no network disclosure path: entries accumulate from ordinary authenticated activity, and an attacker's own requests deposit only their own digest. Exploitation requires independent read access to the audit store, such as database credentials, an audit role, a backup, or a log shipper, at which point the digests support offline password attack against every active account. Whether the material persists depends on the data layer, since raw-term stores keep it verbatim while a SQL store raises Protocol.UndefinedError and drops the entry unless the user resource derives Jason.Encoder. This issue affects ash_authentication: from 4.12.0 before 4.15.0 and from 5.0.0-rc.0 before 5.0.0-rc.2.