BerriAI (LiteLLM) security incidents and vulnerabilities in 2026
BerriAI maintains LiteLLM, an open-source AI gateway that sits between applications and model providers and stores the organisation's provider keys. This page separates LiteLLM's own confirmed release compromises from vulnerabilities in the software, which are not breaches of BerriAI.
Short answer
Was LiteLLM hacked?
Yes, twice, by LiteLLM's own account: malicious litellm 1.82.7 and 1.82.8 were published to PyPI on 24 March 2026, and on 21 October 2026 an attacker who had hijacked its release pipeline published a malicious 1.107.0 to PyPI and its Docker images. Both carried credential-stealing code.
CVE-2026-40217 is different: it is a vulnerability, not a breach of BerriAI. It is a sandbox escape in the LiteLLM proxy's custom-code guardrail test endpoint, fixed in 1.83.11. CISA added it to KEV on 18 November 2026 (due 2 December), and Google Threat Intelligence Group says an actor used it to turn limited inference keys into admin credentials in LLMjacking intrusions.
Timeline at a glance
24 Mar 2026Malicious litellm 1.82.7 and 1.82.8 published to PyPI with a credential stealer. (LiteLLM)Confirmed
8 Apr 2026X41 D-Sec publishes the custom-code guardrail sandbox escape (CVE-2026-40217). (X41 D-Sec)Not a breach
22 Apr 2026LiteLLM 1.83.11 released, the version LiteLLM's advisory names as the fix. (LiteLLM)Not a breach
21 Oct 2026Release pipeline hijacked; malicious 1.107.0 published to PyPI and Docker images. (LiteLLM)Confirmed
16 Nov 2026GTIG reports CVE-2026-40217 exploited in LLMjacking intrusions. (GTIG)Reported
18 Nov 2026CISA adds CVE-2026-40217 to KEV, due 2 December 2026. (CISA)Not a breach
Last updated 20 November 2026. Every entry is dated and linked to its source; claims by attackers are labelled as claims, never stated as fact.
BerriAI (LiteLLM) incident timeline, in detail
Confirmed = acknowledged by BerriAI (LiteLLM), a regulator filing or a court. Reported = credible reporting BerriAI (LiteLLM) has not fully confirmed. Unverified claim = an attacker or leak-site claim.
Nov 2026Not a breachProduct vulnerability (authenticated), known exploited
CVE-2026-40217: sandbox escape in the custom-code guardrail, exploited for LLMjacking
What happened. POST /guardrails/test_custom_code ran submitted Python in a hand-rolled sandbox that could be escaped, giving code execution in the proxy process (root in the default Docker image). LiteLLM lists 1.81.8 up to 1.83.10 as affected and says the fix is in 1.83.11, which uses RestrictedPython. Workaround: block the endpoint at a reverse proxy.
Impact. GTIG says it observed an actor exploit it "to elevate limited-access inference keys ('Virtual Keys') into full administrative credentials", then persist with privileged service accounts, in LLMjacking intrusions (theft and resale of model access). CISA lists it as exploited, due 2 December 2026.
What happened. LiteLLM says an attacker took maintainer GitHub tokens, pointed the release workflow at a lookalike fork and published 1.107.0 to PyPI and the official Docker images on 21 October 2026; it was removed within hours.
Impact. LiteLLM says the release contained malware that steals credentials and runs whenever Python starts. It advises treating affected hosts as compromised, reinstalling pinned to 1.106.2 and rotating every key the proxy held.
Mar 2026ConfirmedSupply chain compromise (package registry)
Malicious litellm 1.82.7 and 1.82.8 on PyPI
What happened. LiteLLM says the attacker bypassed its CI/CD and uploaded the two versions directly to PyPI on 24 March 2026.
Impact. LiteLLM says the versions harvested environment variables, SSH keys, cloud credentials, Kubernetes tokens and database passwords. The official Docker image was not affected, per LiteLLM.
CVE-2026-40217 privileges: LiteLLM's advisory says default configurations need a proxy-admin credential and rates it moderate (CVSS 4.0 5.9); the CVE record scores CVSS 3.1 8.8 with low privileges; X41 used the master key; GTIG says an actor escalated from limited virtual keys.
Fixed version: LiteLLM's advisory metadata says 1.83.10, its text says 1.83.11. X41's advisory (8 April) said no patch was available yet.
The pattern across the record
An AI gateway is a key vault
LiteLLM holds every upstream model-provider key. A compromise of the gateway, or of its releases, exposes all of them, and LLMjacking turns those keys into resold model access and large bills.
Supply chain and vulnerabilities are different risks
The March and October incidents were compromises of LiteLLM's publishing. CVE-2026-40217 is a code flaw that affects everyone running an unpatched version.
What it means for BerriAI (LiteLLM) customers — and what to do
If you run LiteLLM, check which version you run and where it came from, and treat the provider keys it holds as the asset to protect.
1.Upgrade and pin
1.83.11 or later, pinned to a verified release; never 1.82.7, 1.82.8 or 1.107.0.
Its releases were compromised twice in 2026, by LiteLLM's own account: 1.82.7 and 1.82.8 in March and 1.107.0 in October. CVE-2026-40217 is a separate vulnerability, not a breach.
Which LiteLLM version fixes CVE-2026-40217?+
1.83.11 or later, per LiteLLM's advisory. Avoid the malicious 1.82.7, 1.82.8 and 1.107.0.
What is LLMjacking?+
Stealing an organisation's AI model credentials and cloud compute and reselling the access. GTIG says CVE-2026-40217 was used in such intrusions to turn limited keys into admin credentials.
More in the Breach Index
Same sourced, verification-tagged timeline, company by company — compare how peers have fared.
Independent analysis for security-leadership education; not affiliated with or endorsed by BerriAI (LiteLLM). Facts are attributed to the dated sources above, and anything disputed or unconfirmed is labelled as such. See the full Breach Index.