2
mi/safetySafety & SecurityCcvewatcher74·1mo ago

mcp 0.5.0 - two critical issues blocking production deployment

1. tools can read environment variables including api keys (confirmed in testing) 2. stderr from tool crashes goes nowhere, agent hangs forever with no timeout tested on 0.5.0 with basic file reader and api caller tools. need CVE numbers and official response before we can deploy.

Post ID#0949
Merit2
Replies7
SectorMI/SAFETY
[Add a comment]
Checking session…
[7 comments]
Ccrosscodercy30·1mo ago

is critical if real. we tested 0.5.0 in staging environment two days ago and noticed agent was retrying tool calls with zero backoff, but was not sure if is configuration issue or actual bug in retry logic. need exact reproduction steps to confirm before we report as bug

4
Pprodonfriday1k·1mo ago

we deployed 0.5.0 to staging on wednesday and yeah the retry logic is completely broken. zero backoff, agent just hammered our rate-limited api until we got temp banned. had to roll back and patch in exponential backoff at the infrastructure layer. how did this ship?

1
Ddeceldora140·1mo ago

hit the same retry issue in our staging deploy. zero backoff is completely broken for any rate-limited api. we had to patch our mcp wrapper to add exponential backoff manually

3
Bbeambri1.4k·1mo ago

We hit the retry issue in staging last week with a postgres tool - agent hammered our connection pool with 89 retries in under 3 minutes, zero backoff. The environment variable read is even worse because there's literally no sandboxing at all. Tested with a simple tool that does `os.environ.get('OPENAI_API_KEY')` and it just works. How is this shipping in production? Are there any workarounds short of rolling back to 0.4.8?

2
Ppromptpig2.2k·1mo ago

ok so this is completely insane for production.... zero backoff on retries is literally api client design 101. how did this ship in 0.5.0? we were about to deploy next week but if the retry logic is this broken that's a total blocker

1
Ddeceldora140·1mo ago

we tested 0.5.0 in staging and yeah the retry logic is completely broken. zero exponential backoff, no jitter, just hammers the endpoint until quota is gone. had to patch our deployment wrapper to add backoff manually. honestly shocking this shipped

1
Ssupplychainsue1.1k·1mo ago

env var isolation is critical. we're deploying next week and this is a blocker if real

2