AI Brief, 17 November 2026: the keys go in the env file, not the chat
Two official documents published three days apart reach the same conclusion, and between them they close a question this brief put to readers on 30 October. Google Threat Intelligence Group's quarterly AI Threat Tracker, published 16 November as Agentic Threat Activity Crosses a Significant Threshold, reports that in the past quarter GTIG "identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time." The FBI's FLASH-20261113-001, dated 13 November and picked up more widely on the 16th, is a thirteen-page threat overview of a single framework: ARTEX. Both documents name it. The 30 October brief, covering CrowdStrike's report on a China-nexus actor using ARTEX, asked whether the framework would spread beyond the groups then known to use it. It has.
The FBI's figures are the concrete ones. ARTEX has been used since at least August 2026 against organisations across ten sectors, from financial services and government to semiconductors, healthcare and law. Some intrusions progressed "from initial targeting to privileged access in under three hours." A campaign against South Korean financial institutions exfiltrated personal information on approximately 67,000 individuals; a supply-chain compromise of ConnectWise Cloud Backup customers reached more than 2,250 downstream tenants. The developer pulled the GitHub repository on 8 October 2026, after roughly 1,500 to 2,000 stars in ten weeks, but forks, mirrors, archives and a commercial ARTEX PRO keep it in circulation: researchers counted over 1,600 exposed ARTEX consoles in early November.
The detail that should change an engineer's week is narrower. ARTEX ships no model of its own. It integrates external providers, and the FBI reports that on attacker-controlled instances Chinese models are the most prevalent, with the DeepSeek Flash family dominating, attributed to capability, low cost and a "lack of observable guardrails when framed for penetration testing." Consequently the advisory asks defenders to treat outbound connections to LLM API endpoints, unfamiliar model proxies and resellers in egress logs, and "token consumption patterns or API billing anomalies" as indicators of compromise. Your inference bill is now a security signal.
The same day supplied an unplanned counterpoint. OpenCode shipped a product that gives coding agents the user's real browser, with their real logins; hours later a solo tool built on the opposite premise, that the credential never leaves the laptop, reached the Hacker News front page. Monday's argument was about who holds the key.
- GTIG's 16 November tracker reports the first observed case of an LLM framework automating vulnerability discovery
and exploitation end to end, via a multi-agent framework it dubs
exploit_agent. - FBI FLASH-20261113-001 on ARTEX: under three hours to privileged access in some intrusions, 2,250 ConnectWise tenants, roughly 67,000 individuals, 1,600-plus exposed consoles.
- The FBI asks defenders to watch egress to
api.anthropic.com,api.deepseek.com,open.bigmodel.cnandapi.x.ai, plus token-consumption and billing anomalies. - Barracuda's account of a 44-hour intrusion traces it to a cloud key issued to an AI vendor and never revoked.
- OpenCode Browse hands agents a real browser profile with live sessions; lend brokers every credential use behind a local approval.
- MLPerf Endpoints v1.0 arrived on 16 November, the suite's first agentic round, scored in throughput per watt.
Two advisories, one framework, and a new class of indicator
ARTEX is an autonomous multi-agent penetration-testing platform written in Go. The FBI describes it translating high-level goals stated in natural language into machine commands run by specialised agents, "requiring minimal human interaction beyond initial objective setting." Carrying no weights, it brokers whichever provider the operator configures: the advisory lists Anthropic Claude and OpenAI GPT alongside DeepSeek, GLM, Grok and Kimi.
GTIG covers what the FBI document does not: a framework doing vulnerability research rather than exploiting known
flaws. It found exploit_agent on an exposed staging server, set up to run "continuous differential analysis of
popular web application software. It compares code before and after patch announcements, analyzes findings to identify
potential vulnerabilities, drafts exploits, and automatically executes them without human intervention." One figure
shows it reading an access-control check in the Spectra WordPress plugin at version 2.19.25; victim telemetry pointed at
French media, entertainment and education sites.
That target choice is rational, and GTIG explains why: weaponising an already-patched n-day flaw means searching a finite set of source materials, two code versions plus release notes and any public proof-of-concept, where hunting a zero-day means casting a wider net. The effect is lopsided: across January to August 2026 overall disclosures and exploits nearly doubled, while zero-days rose only marginally.
The most telling artifact in either document is a configuration file. GTIG reproduces the environment file that primes
exploit_agent, with comments in Russian:
# LLM API-ключи для авто-эксплуатации (пилот exploit_agent)
# Впиши ключи сам — сюда, не в чат. Обе строки обязательны.
#
# "LLM API keys for auto-exploitation (exploit_agent pilot)"
# "Enter the keys yourself — right here, not in the chat. Both lines are required."
An offensive framework whose setup notes warn the operator not to paste keys into a chat window is a reminder that the economics here run on somebody's API credits. GTIG's third named trend is exactly that: LLMJacking, stealing model credentials from compromised environments and reselling access through underground proxies.
The worked example: 44 hours, one key nobody revoked
The clearest account from the defender's side is Barracuda's, published 11 November by its CISO Arve Kjoelen as "Anatomy of a machine-speed intrusion". Barracuda had issued a cloud access key to a third-party AI vendor and did not retire it when the work ended. The vendor's own development environment was compromised in October, and the key went with it.
The first call with that key came at 04:33 UTC on 1 November, and within minutes the attacker had mapped what it reached and read stored secrets, including AWS access keys and Azure service principal secrets. Each secret opened the next system: cloud accounts, a data analytics system, a source-code access token, Kubernetes clusters. In the early hours of 2 November, inside four hours, short-lived branches pushed to several repositories triggered build workflows that printed build-system secrets to the job log, encoded to defeat log masking. Last recorded activity was 23:31 on 2 November: about ten cloud accounts in roughly 44 hours. Detection was human and incidental, on 4 November, when an engineer noticed unusual activity on their own account, by which time the run had stopped.
Barracuda attributes the tooling, not the actor, to ARTEX, and says plainly that its logging was not detailed enough to rule out access to the credential store behind its XDR product, so it asked affected customers to revoke and recreate integration credentials. For scale, the same post notes that Railway saw a comparable chain in October in which stolen AI-service credentials generated an estimated $740,000 in unauthorised compute, and that at least nine AI platforms disclosed incidents that month.
Two caveats belong on this. GTIG does not establish that exploit_agent caused any specific breach: the server also
held exfiltrated data and post-compromise scripts, but causation is not claimed. And GTIG notes an effect cutting
against the panic reading, that the lower barrier to entry "has fueled a wave of LLM-generated exploits that are
claimed to work but don't, creating additional verification work for defenders."
The same week, agents were handed more credentials, not fewer
Two releases in the window take opposite positions on the same question, hours apart on Hacker News.
OpenCode Browse, an MCP server plus a Chromium extension, gives coding agents "your real browser, with your tabs, your logins and your extensions." The design is deliberate: it uses the existing profile rather than a fresh one, so there are "no keys, no setup," and agent tabs open in background collapsed groups. When a step needs a login, a 2FA prompt or a payment, the agent hands the tab back and resumes when the human clicks Continue, and an Allowed sites list can pin a tab to named domains. It lists OpenCode, Claude Code, Cursor, Codex and Zed as clients, and took 170 points. The page does not say how session data is protected.
lend, by Antonio Mika, starts from the opposite assumption. It runs on the
laptop in front of the existing SSH agent, and a remote VM forwards lend's socket instead of the real one. When the
agent asks to run a command, the policy is evaluated locally, the human approves, the command runs on the laptop, and
only stdout, stderr and the exit code travel back. The default is deny, SSH keys are filtered by fingerprint,
max_signatures caps signing per session, and leases grant a time- or use-limited approval for one exact command.
Secrets are injected at approval time rather than handed over:
{
"default": "deny",
"commands": [{
"executable": "/opt/homebrew/bin/gh",
"args_prefix": ["pr", "view"],
"approval": "biometric:inline",
"secret_reader": "/opt/homebrew/bin/op",
"secrets": { "GH_TOKEN": "op://Development/github/token" }
}]
}
The token never reaches the VM, and the VM cannot choose which references resolve. The author is candid about the limits: an approved command runs unsandboxed as the user, so approving a shell approves everything, and caller details reported by the VM can be spoofed. It reached the front page at 01:18 UTC on 17 November with 63 points, roughly a quarter of the attention Browse got, which is a fair summary of where the field's enthusiasm sits.
Windows 11 records which coding agent you ran
The loudest technical story of the window, at 370 points and 232 comments, is independent reverse engineering. On
16 November Xusheng Li published an analysis of
MissionControlAggregator.dll in Windows 11, reporting that it carries executable-name matchers for coding agents and
developer tools and reports what it finds through Windows diagnostics.
The named slots are specific. Four CLI agents: copilot.exe, claude.exe, codex.exe and agy.exe, with
extensionless equivalents on the WSL path. Five desktop agents: github.exe, claude.exe, codex.exe, cursor.exe
and antigravity.exe. Twelve developer-tool categories cover node.exe, python.exe, gcc.exe, javac.exe and
code.exe among others, some with a publisher check: the Node slot verifies "Node.js Foundation".
What is sent is narrow, and that matters as much as the list. The component subscribes to existing providers,
KernelProcess.AppStateChangeSummary for CLI matches, Win32kTraceLogging.AppInteractivitySummary for desktop ones,
and Microsoft.Windows.Lxss.Manager.Exec for WSL. Those source events carry runtime, time in focus, interaction
counters and resource usage, but the aggregator appears to keep only a bitmap, one bit per category, laid out as
ReportCohorts<12, 4, 5>: twelve tool bits, four CLI-agent bits, five desktop-agent bits. A gcc.exe test set bit
three, giving the value 8. The event is Microsoft.Windows.MissionControlAggregator.Segments.SubCohorts version 0.1.5,
it carries no prompt text and no source code, and its timestamps reflect the collection window rather than usage
duration.
It is marked "isRequired": false, so it travels only with optional diagnostic data, which the US setup path enables by
default, and the behaviour sits behind a feature flag, CLICodingAgentUsage. The author is careful on scope: that flag
was already enabled as a service experiment on the tested virtual machine, while a clean ISO had it off by image
default, so this is not a claim about every installation. The off switch is the ordinary one, under Settings, Privacy
and security, Diagnostics and feedback.
The finding is a measurement, not a leak: what leaves the machine is a nine-bit answer to "which agents does this run." Read commercially rather than as a privacy story, it is a competitive-intelligence feed on agent adoption, taken from the machines of the developers choosing between them.
Also notable
- MLPerf gained an agentic round, and its first headline is about power. MLCommons published Training v6.1 results
on 16 November, and with them MLPerf Endpoints v1.0,
scoring agentic coding and enterprise workflows on throughput
per watt while checking that agents still complete tasks. NVIDIA's debut Vera Rubin NVL72 figures deserve care: on the
Kimi K3 workload, 32 accelerators at concurrency 1,024 drawing 100.89 kW for 41,414.2 tokens/sec, against 48 GB300
accelerators at concurrency 48 drawing 60 kW for 2,210.4 tokens/sec, give roughly
against tokens per second per kilowatt. That is the claimed tenfold-plus generational gain, held at a common floor near 50 tokens/sec/user, but normalised on per-user speed rather than operating point, and time to first token moves the other way, 4,673.9 ms against 2,305.1 ms. The Vera Rubin entries are provisional; the GB300 figures are verified. - The Xena Project asked what "checked in Lean" establishes, reaching Hacker News on 16 November with 293 points.
The post, dated 12 November on the
formalisation-teaching blog, is the most useful follow-up yet to the Fermat's Last Theorem formalisation covered here
on 10 November. It separates four questions a compiling proof does not answer, and the sharpest is statement
faithfulness: of 290 OpenAI theorems with Lean code, "around 38" do not state the paper's main theorem, and one
proves a corollary rather than the headline result. It puts Anthropic's formalisation at about 13 million lines,
needing a 96-core machine with 500 GB of RAM, where OpenAI's 640,000-line Navier-Stokes work compiles on a laptop.
The remedy proposed is a
lake comparatorcheck that the trusted statement matches the untrusted solution and that only standard axioms are used. - California's Attorney General sued OpenAI over agent incidents on 16 November, per coverage citing Politico, reportedly framing autonomous agent behaviour as a product-safety failure across more than 80 incidents; the filing itself is secondhand here. US export controls were extended to remote compute access the same day, closing the cloud-rental route that sat outside rules written around hardware shipment.
- Anthropic told investors a Claude Mythos preview ported its inference stack to AMD GPUs in about three days, work normally measured in months, with "basically no human intervention." Self-reported, with no artifact and no reproduction.
What to watch
- Whether any model provider responds to being named as an indicator of compromise. The useful response is not a statement but a product: abuse telemetry a customer can query, or a path for reporting a reseller-fronted key. The advisory's stated reason attackers prefer DeepSeek is guardrail behaviour under a penetration-testing frame, which is testable by anyone with API access.
- Whether GTIG's
exploit_agentfinding is reproduced by another vendor. One exposed staging server is one data point, and causation for any breach is explicitly not claimed. The figure worth having is how many differential-analysis pipelines are running, counted the way Censys counted exposed ARTEX consoles. - Whether MLPerf Endpoints publishes entries at matched operating points. The first round's headline rests on a comparison between concurrency 1,024 and concurrency 48. A GB300 entry at a throughput-oriented concurrency would separate the hardware generation from the batch size.
- Whether
lake comparatorgets run against the published AI formalisations. The check exists, the Lean code is public, and the result is a count: how many of the 290 theorems state what their papers claim. - Still open: Voxtral 3 weights, day 28, with nothing under
mistralaisince 15 October and an NVFP4 quantisation as the newest artifact there; no independent token count behind the Qwen4-Max efficiency claim, asked on the 16th; no threshold-conditioned fail-open rate from any decision-model vendor, day 39; no independent latency figure for Haru-4B; no Edge0 ablation separating quantisation from post-training, asked on the 13th; and no response from any company named in the Blackwell checkpoint forensics of the 13th.