← All notes G Contributor
Gábor Founder, No Monkey Business
Builds local, source-linked AI systems for law firms. Writes about the infrastructure, access controls and cost truth around them — the parts that decide whether an AI system can be trusted with real work.
2 notes
NMB · FIG.01
LITELLM · SPEND TRACKING VS BUDGET ENFORCEMENT
Tracking is not enforcement.
one request · three tag locations · two outcomes
01
REQUEST PATHS
body tags
metadata
x-litellm-tags header
02
BUDGET GATE
reads the tag representation
the path actually carried
checks key · customer · router
03
PROVIDER
the call that costs money
only when the gate allows it
checked
allowed
reject 429
path the gate never sees
no gate ran
ENFORCED
429 before the provider call
recorded spend does not move
FAIL-OPEN
provider called, spend logged
nothing rejected, dashboard green
Both outcomes look the same on a spend dashboard.
01 Cost control 14 Nov 2026
How spend tracking can keep working while budget enforcement skips a request path, the matrix to test, and the evidence that proves a limit actually rejects.
#litellm #budgets #cost-control #ai-gateway
ANTHROPIC · SEPTEMBER 2026 MISUSE REPORT · THE PATTERN
The attack is a loop. The boundary is four gates.
the loop
minutes, not weeks
01
PROMPT
02
TOOL CALL
03
EFFECT
04
FEEDBACK
files · processes · network · keys
every effect must pass
01
IDENTITY
every agent, tool, MCP server, key —
an owner and a scope
02
POLICY
authorization is argument-aware —
not just “agent can call tool”
03
EGRESS
egress allowlisted and inspected —
stolen code cannot simply leave
04
EVIDENCE
signed trail: prompt → tool → effect →
response → cost
02 Agent security 13 Sept 2026
An analysis of identity, policy, network access, and audit trails in agent systems.
#agent-security #identity #audit #sovereignty