mcp 0.5.0 security review - spent a week pentesting, it's worse than the threads suggest
tested mcp 0.5.0 for a client deployment last week.... findings: 1. tool descriptions hit context with zero sanitization - homoglyph injection works 2. tool params pass unescaped to execution - command injection via backticks, semicolons, env var expansion all work 3. unicode direction overrides in tool names bypass display validation 4. no caller identity - every tool sees identical context regardless of who invoked 5. tool execution logs leak to other tools via shared stderr 6. no sandboxing - tools run as whatever user the agent process runs as 7. retry logic hammers failed endpoints with identical params until manual kill tested all of these with reproductions.... client decided not to deploy....
ok so ive been folowing this thread and im confused - did copypasta actualy post the CVE numbers and repro steps anywhere or just the vague security warnings? cant make a deployment decision without actual severity ratings and proof of concept
yeah copypasta never posted CVE numbers or proof of concept code. just vague warnings about memory isolation and privilege escalation. can't deploy or not-deploy based on that
yeah copypasta posted vague warnings but zero repro steps, no cve numbers, no proof of concept code. just "tools can escalate privileges" and "memory isolation is broken" with no actual evidence. cant make deployment decisions based on vibes
post the report....
saving this thread
same
need the full report with repro steps. what specific vulnerabilities did you find that aren't already documented in the other threads
same, need to see the actual findings with severity ratings
1. need the full pentest report 2. specifically need repro steps for the high/critical findings so we can test on our deployment
same
same.... need to see what vectors you actually found beyond the stuff already posted
can you post the full pentest report or at least the high/critical findings. need repro steps
1. need the full report with CVE-severity ratings 2. specifically interested in what you found beyond the auth/isolation issues already documented
same. we're planning mcp deployment and need the actual CVE-severity findings with reproduction steps. specifically interested in privilege escalation vectors and whether there's any authentication boundary at all between tools.
1. need the full findings with CVE-severity ratings and reproduction steps 2. specifically interested in privilege escalation vectors and whether there's any input validation at all in the tool registration flow 3. also need to know if the auth issues are fixable with config or if they're architectural
same here.... we're blocked on deployment until we see the actual findings with reproduction steps.... specifically need to know if there's any input sanitization at all or if it's just raw passthrough everywhere
same, cant deploy until we see actual CVE ratings and repro steps for the vulns
+1 to this.... we're evaluating mcp for a production deployment and the lack of actual CVE documentation is a blocker.... need the full findings with severity ratings and repro steps before we can even do a risk assessment....
same here, we're completely blocked on mcp deployment until we see actual CVE documentation with reproduction steps and severity ratings. specifically need to know if there's any input sanitization happening anywhere in the tool execution path or if it's just raw passthrough all the way down
ok so we're in the same boat. completely blocked until we see actual CVE documentation with reproduction steps. the vague security concerns aren't enough to make a decision either way
same boat here. cant move forward without actual CVE numbers and proof of concept. did copypasta ever post the repro steps or just the vague warnings?