1. cook.pth imports cook.py, a credential stealer. It collects SSH keys, cloud credentials, CI/CD secrets, AI tool

   tokens, browser passwords, and the LLM provider keys in a LiteLLM proxy database. It encrypts the data and sends it to www.litellm[.]dev.

2. sitecustomize.py starts a background daemon. Every hour, the daemon reads a server address from a GitHub file,

   runs the stealer again, and downloads and runs code from that server.

How the implant got into the release

The pyproject.toml in the 1.107.0 source distribution (sdist) has one extra line in the maturin build section:

pyproject.toml (litellm-1.107.0.tar.gz)

[tool.maturin]
...
python-source = "."
data = "cook"                     # <-- injected
python-packages = ["litellm_cli"] # legitimate (present upstream)

TOML 6 lines

The maturin data = "cook" setting copies a cook/ folder into litellm-1.107.0.data/purelib/ in the wheel. pip installs that folder into site-packages.

The sdist leaves out the cook/ folder, so this line is the only visible change. No public commit in BerriAI/litellm contains data = "cook" or a cook/ folder.

The wheel RECORD lists the three added files:

litellm-1.107.0.dist-info/RECORD

litellm-1.107.0.data/purelib/cook.pth           12 bytes   "import cook"
litellm-1.107.0.data/purelib/cook.py         22252 bytes   exec(zlib.decompress(b64decode(...)))
litellm-1.107.0.data/purelib/sitecustomize.py 25509 bytes  loader/backdoor + embedded stealer

TEXT 4 lines

The release does not match its tag

The tag v1.107.0 points to commit 66c8f5b (06:59 UTC). Its pyproject.toml says version = "1.108.0". SafeDep compared every sdist file with the git trees by blob hash. The sdist matches a later main commit, a8113b0 (08:50 UTC), plus five edited files and the cook/ folder.

Four of the edited files are Cargo.toml files that gain a readme = "README.md" line. The legitimate 1.106.2 and rc2 releases have the same edit. The fifth is pyproject.toml, compared here with a8113b0:

pyproject.toml (a8113b0 → litellm-1.107.0.tar.gz)

-version = "1.108.0"
+version = "1.107.0"
-    "litellm-proxy-extras==0.4.111",
-    "litellm-enterprise==0.1.79",
+    "litellm-proxy-extras==0.4.110",     # same pins as 1.107.0rc2
+    "litellm-enterprise==0.1.78",
+data = "cook"                            # the injection
 ...
-version = "1.108.0"                      # second version field
+version = "1.107.0"

DIFF 11 lines

The edit sets the version back to 1.107.0 and pins the companion packages to the rc2 values. As a result, a dependency diff against rc2 shows nothing unusual. The legitimate 1.106.2 sdist matches its tag commit 49943cd except for the Cargo.toml readme lines.

How the wheels were built

The malicious wheels match the output of the normal BerriAI release pipeline:

• The WHEEL file names maturin 1.9.4, the same version as in the legitimate releases. The implant files have correct RECORD entries, which shows no edits after the build.
• Zip timestamps show separate native builds: macOS at 11:20, Linux at 11:21, and Windows at 11:37 UTC.

• In the Windows wheel, all 2,804 .py files have CRLF line endings, including the three implant files. The Linux

  wheel has none. A git checkout on Windows with line-ending conversion produces this pattern.

No public GitHub Actions run built or published the release. SafeDep checked all 539 runs between 09:30 and 12:30 UTC. At 11:50 UTC, about 10 minutes after the PyPI upload, litellm-dispatcher[bot] published the GitHub release v1.107.0. The same delay occurred for 1.106.2. PyPI has no provenance attestations for 1.107.0, 1.106.2, or rc2.

SafeDep assesses, with moderate to high confidence, that BerriAI’s private release pipeline built these wheels from an attacker-modified git source. The attacker could have used a compromised maintainer or bot credential, a private branch or mirror, or the reference the pipeline builds from. A stolen PyPI token used from the attacker’s own machines fits the evidence less well. Only BerriAI’s pipeline logs can confirm the cause.

Payload 1: the credential stealer in cook.py

At startup, Python runs each line that starts with import in a .pth file in site-packages. cook.pth contains import cook, so the stealer runs in every Python process on the host: the application, pip, IDE language servers, and CI jobs.

cook.py wraps about 61 KB of Python in one layer of exec(zlib.decompress(base64.b64decode('...'))). The decoded code starts with this configuration:

cook.py (decoded)

PC_URL       = 'http://www.litellm.dev/healthcheck'
PC_TOKEN     = 'XWiA8gZNa7NcX2XRoAEb'          # sent as X-Token header
PC_PUBKEY    = '-----BEGIN PUBLIC KEY----- MIICIjANBgkqhkiG9w0BAQEFAAOCAg8AMIICCgKCAgEAvxRBQMVXDXVskKS47haU ...'  # RSA-4096
PC_ONCE_PATH = '/tmp/.pc_once_72129118c59f5636c0bb'
PC_ASYNC     = True                              # runs in a daemon thread
PC_MODULES   = ['sysinfo','netinfo','browser_pw','browser_hist','ssh_keys',
                'shell_hist','cloud_creds','dev_secrets','container_creds',
                'cicd_secrets','ai_devtools','env_vars','db_litellm',
                'cloud_metadata','gpg_keys','tls_privkey']

PYTHON 10 lines

What it collects

How it sends the data

1. It encrypts the JSON results with a random AES-256-CBC key, and encrypts that key with RSA-OAEP and the embedded

   public key.

2. It packs the result into a tar.gz (hdr.json, data.bin, key.bin) and sends it in a POST request to

   www.litellm[.]dev/healthcheck with the header X-Token: XWiA8gZNa7NcX2XRoAEb.

3. After a successful upload, it writes /tmp/.pc_once_72129118c59f5636c0bb so it collects only once per host. It

   also sets PRIVCOLLECT_ACTIVE=1 to avoid a second run in the same process tree.

Payload 2: the backdoor in sitecustomize.py

Python imports sitecustomize at startup, so this file also runs in every Python process.

sitecustomize.py

GITHUB_RAW_URL = "https://raw.githubusercontent.com/letters9813/pythonclient-test/main/domain.txt"
CODE_PATH      = "/codepy"
LOOP_INTERVAL  = 3600
LOCK_FILE      = "/tmp/remote_code_loader.lock"

PYTHON 5 lines

1. It takes an exclusive lock on /tmp/remote_code_loader.lock, so only one daemon runs per host.

2. It calls fork() and setsid(), sends standard input and output to /dev/null, and changes directory to / to

   detach from the parent process.

3. Every hour, it:
   1. reads domain.txt from the attacker’s GitHub repository and takes the first line that looks like a domain,
   2. runs its embedded copy of the stealer, which has no once-per-host marker,
   3. downloads http://<domain>/codepy and runs it with exec().

The attacker can point the daemon at a new server by editing domain.txt. Each infected host runs whatever that server returns as codepy.

At 13:40 UTC, the GitHub account letters9813 and its repository returned 404, about two hours after the release. The backdoor now finds no server address and does nothing. Hosts that ran the daemon before the deletion could have received and run second-stage code. SafeDep could not recover the server address.

Timeline

All times are UTC on 2026-10-21.

Indicators of compromise

Network

Host

File hashes (SHA-256)

The planted files have different hashes on Windows because of the CRLF line endings: