Your Agent's Settings Are Now a Door
What Google Found
On November 16, 2026, Google Threat Intelligence Group (GTIG) published its latest report on how attackers use AI. The title says agentic threat activity has crossed a significant threshold, and the body backs it up: in the past quarter, they identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation "without a human in the loop for the first time."
This is no longer someone asking a chatbot to polish a phishing email. It's several agents splitting the work: one reads the vendor's patch, another works out which hole it closed, another writes the exploit, another runs it. Google calls it a paradigm shift towards "intrusions occurring at machine speed."
The numbers point the same way. Between January and August 2026, overall disclosures and exploits nearly doubled. Zero-days — attacked before a patch exists — went from a 2025 average of 8 a month to 11 a month in 2026, then spiked to 22 in August. And in the cases where attackers hijacked cloud AI accounts, Mandiant saw single-day fraudulent charges above $100,000 and, in the worst cases, $1 million in cumulative abuse before anyone stopped it.
The Line That Stopped Us
Halfway through the report there's a paragraph about a suspected North Korea–nexus cluster Google tracks as MIDNIGHT NEPTUNE. It built fake recruiter profiles with AI-generated photos and steered targets into fake meetings. Once inside a company, it went into the internal code repositories and modified .claude/settings.json files, so that when someone started Claude Code, it automatically downloaded a backdoor called SOMBERMEME.
Claude Code is the tool we build with every day. That file is the one that tells the agent what to do the moment it starts.
The same day we read the report, we had added a startup instruction to that very file ourselves: a reminder that, every 45 days, tells the agent to clean up its own rules. Same mechanism, different intent. That's what makes this finding uncomfortable: there is nothing exotic about the attack. It uses a legitimate feature — the one all of us use to make agents work better — in a file almost nobody opens twice. Google sums it up as attackers "exploiting native AI agent execution hooks as enterprise access vectors."
What Didn't Change
The most useful sentence in the report is the least dramatic one: "the mechanics of the intrusions remain the same, such as using compromised credentials and targeting unpatched systems." What changed is the speed and scale at which those known weaknesses get exploited.
That's half good news. Attack agents didn't invent new magic; they exploit the old carelessness faster. And for now, Google notes, "they currently trade stealth for velocity": they generate forensic traffic, rely on recognizable identifiers and leave code-level traces. They can be caught, if someone is looking.
There's a second effect worth naming. The same report describes a wave of AI-generated exploits "that are claimed to work but don't," adding verification work for defenders. Unchecked model output is noise on the attacker's side too.
What We Changed Ourselves
When we reread our own history through this report, the conclusion was uncomfortable and useful: so far, every security failure we've had came from our own speed, not from an attacker. A new table born without a lock. Client data that ended up where it shouldn't. A key configured wrong. Exactly what an attack agent finds first, because it's what it looks for: credentials and carelessness.
So this month we set an information security methodology based on ISO/IEC 27001 for every Menteorama™ system. We are not certified — certification comes from an outside body, and we won't claim what isn't true. What we did was choose where to start and put a date on it. It isn't 93 controls on paper: it's five things that, by December 15, have to be running, not written down:
Master keys first. The accounts that open everything else require a physical or biometric key. If attackers automate credential theft, a credential alone has to stop being enough.
The line between clients checks itself. A test that runs on every change tries to read one client's data from another client's session. If it succeeds, the change doesn't ship.
Agent configuration is code. The files that tell an agent what to do at startup are reviewed like any production change, live in version control, and only run what we wrote. A startup instruction nobody reviewed is an open door.
What an agent reads is not an order. A web page, an email or a repository can carry hidden instructions. For our agents, all of that is information; orders only come from the person they work with.
Locks that enforce themselves. A written rule protects nothing if it depends on someone remembering it. If the attack runs at machine speed, the defense can't be a document: it has to be a test, a block or an automatic check that fails loudly.
Speed Is Answered With Structure
Google's report closes with what it does as a platform: behavioral detection, hardening guides, models with safeguards. That's the labs' part. Those of us who build with agents own the other part: deciding what the agent can do, with which keys, and who checks.
It's the same division of labor we believe in for everything else. The model brings capability. The method decides where it stops. When attackers have agents too, that structure stops being good practice and becomes the defense.
Created with human intelligence. The agent executes; the method closes the doors.