BerriAI LiteLLM: attackers can get in without a password
OFFICIAL NAME: BerriAI LiteLLM Improper Authentication Vulnerability
WHAT IT IS · AI TOOLSIt connects to AI models and often holds the keys to paid services and data.
DO THIS FIRSTUpdate it and change any API keys stored in it. Not sure if you use it? Send this page to your IT provider or security lead and ask.
TECHNICAL SUMMARY BerriAI LiteLLM contains an improper authentication vulnerability in the MCP Streamable HTTP endpoint that could allow an unauthenticated attacker to establish an authenticated MCP session using an arbitrary Bearer token.
WHY THIS VERDICT Attacks are confirmed. The vulnerability is 116 days old, so current versions are being attacked.
- HIGHEST CONFIRMED RUNG
- EXPLOITED
- CISA KEV
- LISTED
- EXPLOIT CODE
- PROOF OF CONCEPT
How far has it climbed?
01 TALK
▨ REPORTED, NOT CONFIRMEDNEWEST EVIDENCE · 1 SOURCECVE-2026-59822 discussed publicly (CIRCL sighting).
02 EXPOSED
□ NOT TRACKED YETNO SOURCE FOR THIS STEP YETWe have no source that counts exposed systems yet. Internet scan data needs a partnership we do not have. No evidence here does not mean nothing is exposed.
03 TRIED
■ CONFIRMEDNEWEST EVIDENCE · 2 INDEPENDENT SOURCESAttacks that worked are confirmed (step 4), so attempts count as confirmed too.
04 EXPLOITED
■ CONFIRMEDNEWEST EVIDENCE · 2 INDEPENDENT SOURCESBerriAI LiteLLM contains an improper authentication vulnerability in the MCP Streamable HTTP endpoint that could allow an unauthenticated attacker to establish an authenticated MCP session using an arbitrary Bearer token. CISA Vulnrichment marks CVE-2026-59822 as actively exploited.
05 DAMAGE
□ NOT TRACKED YETNO SOURCE FOR THIS STEP YETBreach notices and filings rarely name the flaw that was used, so we cannot link them to this event yet. No evidence here does not mean no harm.
When each thing happened
- DISCLOSED
- EXPLOIT CODE PUBLIC
- FIRST ATTACK REPORT
- CISA KEV LISTING
- US FEDERAL DEADLINE
THE FIX
CISA's instruction: apply the vendor's fix, or stop using the product if no fix exists. The deadline CISA set for US federal agencies, Sep 16, 2026, has passed. It is a US government deadline; use it as a guide.
Questions to ask this week
- Do we run BerriAI LiteLLM? If yes, is it patched or cut off from the internet?
- Who owns fixing it, and by what date?
- Have we checked for signs of attack since the first report, not just since the patch?
How it unfolded
- Attack activity. JUL 8, 2026CISA Vulnrichment marks CVE-2026-59822 as actively exploited
- Attack activity. SEP 1, 2026BerriAI LiteLLM Improper Authentication Vulnerability
- SEP 2, 20261 public GitHub repository claims a proof of concept for CVE-2026-59822
- OCT 26, 2026CVE-2026-59822 discussed publicly (CIRCL sighting)
- OCT 26, 2026Public proof of concept for CVE-2026-59822 (CIRCL sighting)
SOURCES
- [1] CIRCL Vulnerability-Lookup sightings: CVE-2026-59822 discussed publicly (CIRCL sighting), OCT 26
- [2] PoC-in-GitHub: 1 public GitHub repository claims a proof of concept for CVE-2026-59822, SEP 2
- [3] CISA Known Exploited Vulnerabilities: BerriAI LiteLLM Improper Authentication Vulnerability, SEP 1
- [4] CVE List V5 and CISA Vulnrichment: CISA Vulnrichment marks CVE-2026-59822 as actively exploited, JUL 8