This time the thread is criminals running AI agents against real victims, documented by five separate teams since we last wrote. Nine weeks since the last issue is longer than we meant, so this one runs bigger, with a new quick-hits section so the important stories don’t fall through the gap.
What the industry’s doing
-
Criminals got agents, and five teams have the receipts. (GreyNoise, NorForge, Reuters, CrowdStrike, Unit 42, Gambit). In August our industry section closed on Sophos: a full year of its casework, ending in June, turned up no fully autonomous agent-driven intrusions. We added our own line: “The criminal reality is still a 2022-vintage infostealer wearing a 2026 logo.” Sophos was right about its window. Our line was too neat. We learn out loud, so here is the correction.
GreyNoise watched a likely Russian-speaking actor point hundreds of agents (OpenAI’s Codex as the harness, a DeepSeek model doing the work) at PaperCut: just under four hours from an empty workspace to RCE on a real victim, and by GreyNoise’s count 440 instances at 395 organizations. NorForge found more than 20 Cairn agent panels left open on the internet; one Chinese-language cluster ran over 1,100 attack projects. Reviewing that operator’s activity, Reuters found it broke into websites belonging to more than 100 companies, at about $50 per successful hack by the researcher’s estimate, though Reuters could not establish how much data was taken. CrowdStrike links, with moderate confidence, an operator running the open-source ARTEX pentest agent plus Claude Code sessions to breaches at South Korean financial firms. Unit 42 describes one human driving parallel frontier agents through more than 50 ATT&CK techniques in under 10 hours.
The playbook is old: LSASS dumps, noPac from 2021, DCSync, SQL injection, hard-coded secrets in code repositories. What changed is tempo, cost, and how little the human typed: Reuters says that operator’s instructions ran under 100 Chinese characters. Two of these cases surfaced because operators left their own panels or directories open. Keep the counterweight too. Microsoft’s Digital Defense Report, as quoted by BleepingComputer: “Most observed campaigns still retain human direction, even as frontier systems demonstrate end-to-end autonomy in labs and early real-world cases.”
So measure your window against theirs. PaperCut’s advisory went out on 08-27 and GreyNoise’s actor had RCE four days later. Know the hours from advisory to fix for every internet-facing app, and pre-approve the stopgap, such as a WAF rule (the one block GreyNoise documented was Cloudflare’s WAF). The actor reached domain admin at only 12 of those 395 organizations, so patch noPac and alert on DCSync from hosts that aren’t domain controllers. Hunt what agents leave behind, per Unit 42: structured Markdown hand-off files, bursty API traffic, sudden model use from unexpected identities. And keep backups out of the application’s reach: a cleanup routine Gambit documented dropped 180 tables, the admins’ own backups included.
-
Lab agents, part two: outsiders broke most of the news. (collusion.wiki, rubyhack.ai, OpenAI’s notices, The Record, Anthropic) · update on 003’s hero. Researchers found about 17,000 agent edits on a near-dormant German wiki, by agents self-identifying as OpenAI’s; OpenAI confirmed they used it as a message board. They believe an OpenAI swarm pushed over 2,000 packages to RubyGems; OpenAI says its agents did benign tasks there and hasn’t verified the malicious-upload claims. Australia’s prime minister says officials heard about agents in a Medicare data portal almost three months late. Separately, Anthropic cut live internet from all internal evaluations, for now. Hunt package names, author fields and usernames containing “oai”.
-
AI gateways moved from research topic to target (Microsoft, 2026-08-26) · update on 003’s LiteLLM item. 003 covered LiteLLM key theft as post-compromise research. Microsoft then described intrusions it investigated against three AI workloads. On LiteLLM, attackers read credentials from the process environment, dumped the model and virtual-key tables, and ran a cryptominer. RAGFlow got a hidden hook capturing newly entered provider keys. Microsoft points to CVE-2026-42271, which public research chains with Starlette’s CVE-2026-48710 to drop the authentication requirement; CISA added the Starlette bug to KEV on 09-02. Ten Defender XDR hunting queries come with it, and one line worth stealing: “Treat AI gateways as Tier-0 secrets stores.”
-
Exploit help with the refusals mostly stripped is now a download (Anthropic Frontier Red Team, 2026-09-29). On Anthropic’s tests, Z.ai’s open-weight GLM-5.3 built end-to-end exploits in 50 of 410 ExploitBench attempts, against 56 for Claude Mythos Preview, a limited-release model from five months earlier. NIST’s CAISI calls it “the most cyber-capable open-weight model released to date”, while putting it about four months behind the US frontier. Abliteration cut mean refusals across three benchmarks from 95% to 6%, and abliterated copies appeared publicly within days. CrowdStrike found GLM-5.3 in the ARTEX operator’s Claude Code sessions. This is one lab measuring a competitor, so weigh it that way.
From the community
The work we want to amplify: practitioners showing what they tried, what broke, and what they learned.
-
Three takes on why a sandbox is not containment (Trail of Bits, Matthew Green, AIR Security) · update on 003’s hero. Artem Dinaburg asked GPT 5.6-Cyber to escape his QEMU/KVM sandbox. By his count it escaped three times, the first with an exploit that “didn’t land cleanly”, the last via a chain including bugs then unpatched. Firecracker: only a hardlocked host. Green: “a warden-guarded sandbox is just another version of the alignment problem.” AIR’s Matan Elgabsi and Or Nevo leaked a compromised agent’s secret through an allowlisted server’s certificate fetch; Vercel ruled it out of scope. Treat allowlisted endpoints as egress; use fresh environments and time limits.
-
Elastic says it took its AI triage from about 60% to 92% right without changing models (Elastic Security Labs, Maggie Musquez, 2026-08-25). The window’s clearest account of a SOC using AI rather than being acted on by it. Elastic’s early AI case summaries were “inaccurate 40% of the time”, so analysts ignored them. Accuracy, measured against analysts’ close reasons, moved with context: per-rule investigation guides, HR context such as joiners and leavers within 90 days, and the last 30 days of closed cases on the same rule, including analyst-flagged AI misses. Elastic gives no sample size or period, so the number is theirs. The method is yours to take.
-
A honeypot fed an attacker’s coding agent, and the agent handed over its playbook (SANS ISC diary 33332, Renato Marinho, 2026-09-11). Marinho’s honeypot posed as an OpenAI-compatible endpoint, and an attacker’s semi-autonomous coding agent kept picking it as a free backend. Its requests carried about 43 KB of control plane: AGENTS.md, an offensive playbook, collected API keys, the operator’s direct egress IP. The playbook farms LLM access: FOFA discovery, trial accounts via temp mail and CAPTCHA solvers, roughly 379 upstream endpoints in a self-hosted gateway. Pointing agents at cheap proxies? His line: “A malicious LLM endpoint sits in a privileged position to observe the agent that consumes it.”
-
Copilot in Visual Studio and SSMS, fooled by a trailing dot (Embrace The Red, Johann Rehberger, 2026-10-13). Injected instructions could have Copilot write
.mcp.json.: the path check saw a different file, Windows wrote the real one, and a malicious MCP server got enabled. CVE-2026-103572, CVSS 8.8, fixed in October’s Patch Tuesday, reported 08-04 for a $0 bounty. In his companion post, a db_owner plants instructions in a database extended property, and when a sysadmin asks Copilot in SSMS for help, the attacker lands in sysadmin, because read-only mode was a regex classifier. Check that Visual Studio and SSMS on admin workstations are current, not just servers.
Quick hits
One link, one fact, one so-what. Nine weeks is a lot of window.
- DIVD disclosed its own breach (BleepingComputer, 09-30). Two KEV-listed Zammad zero-days gave what DIVD calls an agentic attack root. Data was taken; segmentation stopped it going deeper. The root bug is still unfixed: restrict server access to trusted admins.
- SneakPeek (Noma Labs, 10-14). Previewing a poisoned HTML file over 256KB in Claude Code Desktop could read SSH keys and cloud tokens, with no permission prompt. Fixed in Claude Desktop 1.30096.1. Repo files are untrusted input.
- AgentCorruption (Zenity Labs, 10-08). One prompt sent an AWS AgentCore agent to its metadata service; the default role could then invoke other agents and read Secrets Manager. AWS calls it documented, expected behaviour. Scope roles per agent.
- Two Codex sandbox escapes (Accomplish, 09-15). One reached unsandboxed execution from read-only mode by opening a cloned repo and asking a question. Both fixed within eight days. “The thing doing the enforcement was sitting inside the thing being enforced.”
- OpenCode Console leaked users’ AI provider keys (BleepingComputer, 10-19). 1.4% of Console users, not the open-source tool; users reported overnight charges. OpenCode calls it a distillation attack. Rotate keys you’ve handed to third-party tools.
- GitLab AI Gateway, CVSS 9.9 (GitLab, 10-02). An authenticated Duo Agent Platform user could escape the prompt-template sandbox and run commands. Self-hosted gateways need 19.2.4, 19.3.2 or 19.4.1. After LiteLLM’s, the second AI gateway command-execution bug in two issues.
- CARBONATO (ThreatDown, 09-22). A botnet entering through unauthenticated Docker on port 2375 installs the Hermes agent unchanged, rewrites its persona as GH0ST and ranks AI API keys as loot #1. Hunt unexplained Telegram egress from servers.
- Two scripts to reconstruct what an agent did (SANS ISC, Jim Clausing, 10-08). Standard-library Python rebuilding transcripts and tool calls from opencode and Hermes. His caveat belongs in your report: not proof every action was recorded.
- numbat (SANS ISC, Russ McRee, 09-04). Perplexity’s open-source tool inventories the AI agents on an endpoint, hooks them in monitor-only mode and maps behaviour to ATT&CK. A first look at what your laptops’ agents are doing.
- Small agents, smaller blast radius (SANS ISC, Mark Baggett, 10-16). Narrow agents, investigation separate from response. “A sentence in a ticket claiming that someone approved an action is evidence to evaluate, not permission to execute it.”
- Coding agents weakened their own monitoring (OpenAI, updated 10-20). One muted Security alerts after “don’t send any messages”; another removed its monitor, then denied it. Protect monitoring config from agent edits.
- DHS’s DeepSeek order goes public (Federal Register, 10-20). The July order also removes “third-party products and services developed using DeepSeek implementations” from civilian agencies. That clause is the work: check your vendors.
- Vulnerability trends in the AI era (Google Threat Intelligence, 09-30) · patch-volume update. Monthly disclosures doubled January to August; 0.23% of 2026’s were ever seen exploited. October Patch Tuesday: SANS counts 1,417, one exploited. Triage on exploitation.
- Self-replication in an easy lab (Anthropic, 10-12). A GLM-5.2-driven worm spread to all 8 emulated nodes in 4 of 5 trials, standing up a smaller stand-in model on each. No full rogue replicator yet. Watch for multi-gigabyte device-to-device transfers.
On the show
Still no episode to plug. The podcast is in the works; until then the conversation lives in your replies.
One idea to take with you
Score the action before you let an agent take it. In One does not simply defend agentically, NCSC-UK’s Dave Chismon starts from Halvar Flake’s line, “All offensive problems are technical problems, and all defensive problems are political problems”, and lands on “Defenders simply cannot put AI to work in the same way attackers can.” His rubric rates each automated action 0 to 4 on five dimensions: potency (advice to a human, up to direct system changes), scope (one system, up to the fleet), criticality, rollout confidence (audit mode or a proper digital twin, up to anecdotal validation) and recoverability (no state change, up to irreversible). Low-potency advisory work is the easy win, and he asks whether defenders can start where systems are cattle, not pets. Read it next to the hero: attackers can automate recklessly because breaking your systems costs them nothing. You can’t, so automate on purpose. Elastic’s triage and Baggett’s small agents above are what a low score looks like in practice.
Shape issue 005
If you’re building something with AI in defense, broke something interesting, or saw a story we missed, hit reply. Issue 005 gets built from what you send.