Archive
Format: Security brief
-
REVIEW
Starlette Host-header flaw could undermine path-based security checks
CVE-2026-48710 matters where application authorization relies on a reconstructed URL path that differs from actual routing.
-
REVIEW
LiteLLM guardrail creation paths lacked the checks applied in testing
CVE-2026-59821 concerned custom Python guardrails and administrative access, with additional risk when no master key was configured.
-
REVIEW
LiteLLM Host-header authentication bypass depended on upstream request handling
CVE-2026-49468 affected certain proxy deployments; the advisory says LiteLLM Cloud was not affected.
-
REVIEW
LiteLLM flaw chain linked low-privilege keys to administration and code execution
The reported chain combined three separate authorization and execution weaknesses, rather than one generic Claude vulnerability.
-
REVIEW
LightGlue loading flaw could override a caller’s remote-code trust choice
CVE-2026-5241 involved nested configuration loading in Transformers, with a fix listed in version 5.5.0.
-
REVIEW
Transformers configuration flaw could load attacker-controlled generation code
CVE-2026-4372 involved an internal attention-implementation setting in untrusted model configuration.
-
REVIEW
Vercel traced April 2026 intrusion to a third-party AI tool and employee account
Vercel’s bulletin describes access through Context.ai and Google Workspace, while distinguishing the incident from npm package compromise.
-
REVIEW
Transformers Trainer checkpoint flaw depended on the PyTorch version in use
CVE-2026-1839 involved unsafe loading of RNG state from an attacker-controlled checkpoint, rather than a generic PyTorch service breach.
-
REVIEW
LiteLLM configuration endpoint lacked an administrator-role check
CVE-2026-35029 could let an authenticated user change privileged proxy configuration; version 1.83.0 added the required check.
-
REVIEW
Claude Code packaging error exposed source files, according to Anthropic
The company described an accidental source-map release and said customer data and credentials were not exposed.
-
REVIEW
Mercor confirmed a security incident linked to the LiteLLM supply-chain compromise
The company confirmation and an extortion group’s data-theft claims are separate pieces of evidence.
-
REVIEW
Malicious LiteLLM releases turned an AI gateway dependency into a credential-theft risk
TrendAI’s analysis identifies trojanized package versions 1.82.7 and 1.82.8 in the March 2026 supply-chain incident.