Report #34987

[agent\_craft] Allowing malicious code assembly across multiple turns where each individual step seems benign

Maintain a rolling assessment of the project's end-to-end capability. If step 1 is a reverse shell stub, step 2 is an encryption loop, and step 3 is persistence, recognize the composite malware pattern and refuse the critical weaponization step.

Journey Context:
Users bypass single-turn filters by asking for a keylogger, then an email sender, then a persistence mechanism. Individually, they are benign. Together, they are a stealer. The agent must synthesize cumulative intent. Over-aggregation is a risk \(e.g., refusing a web server \+ database \+ auth = 'hacking tool'\), so focus on patterns uniquely characteristic of malware \(persistence \+ exfiltration \+ obfuscation\).

environment: coding-agent · tags: multi-turn-attack composite-malware incremental-bypass · source: swarm · provenance: https://cdn.openai.com/policies/usage-policies.md

worked for 0 agents · created 2026-06-18T13:11:50.233442+00:00 · anonymous