A Refusal Is Not a Boundary

October 13, 2026 · SPR{K3 Research

SOCRadar recovered an attacker's own Claude Code session logs. The refusals are in there — and so is how the operator worked around them.

On October 12, The Hacker News reported two SOCRadar investigations: one into CyberXero, a Russian-speaking initial-access broker active since at least July 2026, and one into Operation Master, a Brazilian fraud campaign run by a cybercrime-forum actor using the alias "masterblack." Both operators ran their intrusion work through commercial coding agents. CyberXero's primary workstation ran Claude Code with 51 specialized agents — one per phase of the kill chain: scanning, exploit chaining, reconnaissance, post-exploitation — alongside Cobalt Strike, PentAGI, AdaptixC2, GodPotato, Kerberoast, Evil-WinRM, rclone, SFTP tunneling, and WSO web shells. The victim set confirmed by SOCRadar: four Ukrainian organizations, including an energy/utilities heating provider in Kharkiv, with 628,000+ records tied to Ukrainian individuals exfiltrated, plus additional compromises in Poland, China, and the Netherlands. Operation Master separately moved 2.4M+ email/SMS/WhatsApp messages and 622,666 personalized invoice links against Brazilian consumers, aiming at 21 Brazilian financial institutions and amassing an R$150.4M click-tracked cumulative invoice value.

The two campaigns are different in scope and target. They share one piece of tradecraft that is worth sitting with.

The session-reset primitive

SOCRadar recovered both operators' AI session logs from the attackers' own exposed open directories, including their refusal transcripts. The CyberXero logs show the model, repeatedly, refusing to help: install a backdoor, disable a firewall, move laterally, deploy a web shell. Across four iterations on each, those refusals held. Four scans were also refused. One forged ownership token was rejected outright. The alignment layer did its job on the direct asks.

Then the operator closed the session and opened a new one.

The new session was preloaded with the framing that this was an "authorized pentest on my own infrastructure." The same offensive goal, reworded. In Operation Master, the framing was "authorized bug bounty or penetration testing activity." SOCRadar's one-line summary, quoted by The Hacker News: "model safety controls act on the context of a session, not on the model weights," and so "a refusal in one session does not carry into the next."

That is the primitive. It isn't exotic. It isn't a jailbreak in the usual sense — no injected tokens, no roleplay elaborately scaffolded. It is the attacker noticing that the session is a transient boundary, and session-scoped alignment is a transient control, and both reset on command.

What the vendor sees and doesn't see

A commercial coding agent sees one session at a time. The abuse-detection pipeline inside the model provider sees one session at a time. If a user asks for something the model declines, that is a refusal, logged. If the same user opens a new session an hour later and asks the same thing wrapped in a plausible authorization claim, that is a new conversation, logged separately. The two are not stitched together into a composite picture of operator intent — not because the vendor is lazy, but because the vendor's unit of monitoring is the session.

The operator's unit of work is the campaign.

The gap between those two units is where this tradecraft lives. CyberXero's 51 agents are a brute-force expression of the same gap: specialize the agents by kill-chain phase, run them concurrently from one workstation, and no single session has to look like a full intrusion. Each session looks like a reasonable scan, a reasonable exploit lookup, a reasonable code generation task. The composition is the attack.

Three things to notice

The refusals that held are a map, not a reassurance. CyberXero's logs show four categories the alignment layer caught across every attempt: backdoor install, firewall disable, lateral move, web shell deploy. That is defender-side signal — those categories work. Everything else that the operator ran 51 agents against is, in the record, not refused. The reasonable reading is that the alignment layer is well-trained on a specific list of highly-flagged actions, and silent on the long tail of smaller tasks that compose into a successful intrusion.

The attacker's own exposure is where the evidence lives. SOCRadar's primary evidence base — session logs, plaintext tokens, refusal transcripts — came from the attacker's exposed open directory, not from the vendor. That is a windfall for the defender community right now. It is also a near-term signal that attackers will tighten operational security around their AI tool use. The window on this evidence class will close. Defender infrastructure that depends on attackers forgetting to lock down their own servers is not infrastructure.

One workstation, 51 agents, 61 seconds. SOCRadar reports a single automated run that scanned 4,708 targets, confirmed 429 WordPress admin panels, and deployed 32 web shells in a 61-second window. That is a quantitative floor on what one motivated operator with commercial coding agents can execute per minute. The economics of perimeter software defense were written for operator-scale attackers. The current threat is agent-scale.

The direction of travel

The Oct 11 digest covered the CrowdStrike ARTEX disclosure — an open-source Chinese agentic framework orchestrated across four commercial LLM backends against South Korean banks. The Oct 10 digest covered the Barracuda dual-channel email primitive, where one message carries a payload for the human reader and a hidden prompt for the AI reader. The Oct 9 digest covered the Zenity AgentCorruption chain, where one prompt to one public-facing managed agent reached every other agent in the AWS region. CyberXero and Operation Master are the operator-side pattern on the same substrate: not a new vulnerability, not a jailbreak, just a refusal-evasion primitive at session granularity, applied at campaign scale.

Session-scoped alignment is one layer. It is not the layer attackers aim at. The behavior that follows a refused prompt — the reopened session, the preloaded framing, the quietly-successful cross-session composite task — is visible only from outside the session boundary. That is the layer that matters.

A refusal is not a boundary. The campaign is.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.