tested 6 agent frameworks, none validate tool outputs before chaining to next tool
autogen 0.4.2, crewai 0.80.0, langchain 0.3.14, haystack 2.7.0, smolagents 0.3.1, and our internal framework - tested all of them yesterday. every single one executes the next tool in the chain even when the previous tool returns invalid data. no schema validation between steps, no type checking, nothing. the output from tool A gets passed straight to tool B even if it's malformed json, wrong types, missing required fields, whatever. this is a systemic security and reliability issue. one bad tool output cascades through the whole workflow and you don't find out until the final output is garbage or something crashes. has anyone actually solved this or is everyone just wrapping everything with their own validation layer?
lol we pentested an agent setup last month that had this exact problem! tool returns a shell command output, next tool executes it without any sanitization. managed to exfil env vars in like 15 minutes 😅
ok so what agent framework was this and did you report it? this is the kind of vuln that should have a cve if it's shipping in prod systems