C2YA - Inside the Adversary's Inbox

The talk opens with a simple question: if coding agents can find bugs in ordinary software, what do they find in the tools attackers use? The speaker pointed LLM-assisted research at five open-source command-and-control frameworks, Havoc, Mythic, Sliver, Covenant and AdaptixC2, several of which have appeared in operations by state-linked and criminal groups. Over about six months the work produced roughly 250 flaws, many reachable from the internet without prior access, each scored by how useful it is to a defender rather than how much damage it does to a target.

The middle covers the most useful bug classes. Havoc reused an AES-CTR keystream for each agent, so XORed ciphertexts leak plaintext through crib dragging, and its unauthenticated WebSocket hands out the session key. Adaptix embedded its RC4 key in every implant, Sliver kept key material in unencrypted SQLite, and Covenant let a low-privilege user read password hashes and escalate to code execution. A live demo eavesdropped on a Havoc team server knowing only its address. A survey of 35 exposed servers, 13 of them live, found a Havoc campaign against French victims that used layered obfuscation and unprotected cloud buckets.

The closing point is that Havoc was archived in February 2026, so its bugs will not be patched and defenders must carry that load. Taking down a C2 server is rarely enough, because operators leave backup access. Practitioners should treat adversary tooling as software worth auditing, both for fingerprinting and traffic decryption and for recognising that attackers may exploit the same flaws.