cd /news/ai-safety/the-productivity-gains-showed-up-on-… · home › topics › ai-safety › article
[ARTICLE · art-171127] src=fromtheterminal.substack.com ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

The Productivity Gains Showed Up on Offense First

The FBI issued a flash advisory warning that intruders are using Artex, an open-source autonomous AI penetration-testing platform, to breach networks in under three hours, including attacks that took personal data on at least 68,000 people from South Korean financial institutions and reached Microsoft 365 environments of at least 2,250 ConnectWise customers. One breached security vendor's chief said defenders who rely on spotting malicious activity and then responding "will often arrive after the attacker has finished." The developer has since taken Artex closed-source, but the old version still circulates and researchers found hundreds of exposed instances online without passwords.

read6 min views1 publishedNov 17, 2026
The Productivity Gains Showed Up on Offense First
Image: Fromtheterminal (auto-discovered)

AI is compressing the software loop at both ends — building and breaking — and the two ends are not compressing at the same rate.

1. The Attackers Are Running the Same Playbook You Are. They Just Shipped It First. #

The FBI put out a flash advisory this week warning that intruders are running autonomous AI penetration-testing platforms against real targets — financial services, government, defense, telecom, law firms. The tool it names is Artex: open source, pointed at whatever model you like. In some cases it carried an operation from first scan to active compromise in under three hours. It took personal data on at least 68,000 people from South Korean financial institutions, and through the IT vendor ConnectWise it reached the Microsoft 365 environments of at least 2,250 customers.

The tradecraft is almost insultingly ordinary. Artex is noisy enough to trip a competent alerting stack, and it goes after the dullest flaws available: weak configurations, missing multi-factor, exposed or stolen credentials. One breached security vendor traced its incident to a credential it never revoked after ending a vendor engagement; its logs couldn't rule out what the intruders reached, so it rotated every customer credential it held. Nothing in that chain needed a model. What the model changed was the clock, and that vendor's security chief said the quiet part out loud: defenders who rely on spotting malicious activity and then responding will often arrive after the attacker has finished.

Two details stick. The developer took Artex closed-source after the attacks and the old version still circulates — there is no un-ship button. And researchers found hundreds of Artex instances exposed on the internet with no password, including the one used against South Korea: offensive automation is software too, built just as carelessly as everything else.

Why it matters:

  • For ICs: Dead credentials are now scanned and exploited within hours, not months. Revocation on vendor offboarding stopped being hygiene and became a control.
  • For leaders: Detection budgets written against human attackers are obsolete. If your plan assumes a day of dwell time before damage, model three hours and see what survives.
  • For founders: Detect-and-respond is a harder sell now. The buyable story is prevention, blast-radius reduction and automatic revocation.

2. The Code Still Works. Nobody Can Say Why Anymore. #

A much-discussed essay this week argues that 2026's run of outages and breaches isn't bad luck but one failure wearing different clothes. Its exhibits: the multi-day Microsoft outage, where a configuration fault cascaded until no service could be restored long enough to carry traffic; PHP2Shell, an arbitrary-code-execution exploit across the PHP runtime found with AI tooling, resting on memory bookkeeping bugs nobody had noticed in a decade; and the Shai Hulud attack, where a CI misconfiguration let forks push code upstream and install scripts lifted the tokens. The shape is shared: a complex system of low-quality parts where one error compromises everything.

The argument about AI is sharper than the usual one. The complaint isn't that model output is unreliable slop — the essay concedes it's largely good at what you ask. The claim is that it's good locally and incoherent globally, a condition it borrows a name for: decoherence, the accumulation of locally valid but mutually incompatible models, until the cost of change is dominated by translating between them and preventing side effects. Decoherence grows with a system's rate of change — so the thing we bought AI for is the thing manufacturing it.

Strip out the doom and a testable claim remains: time-to-ship is falling and time-to-understand is not falling with it. That gap shows up in reopened incidents, review cycles on agent-authored changes, and how long a new engineer needs to answer "why is this here." It is also the gap the first story exploits: PHP2Shell rested on flaws too complex for a human to spot, which is why offense found them first.

Why it matters:

  • For ICs: Architectural coherence is the scarce skill now, not throughput. Being the person who can say why the system is shaped this way is a defensible position.
  • For leaders: Track time-to-understand next to velocity. If one improves while the other degrades, you're accruing an unlogged liability.
  • For founders: Every agent-built greenfield codebase is a decoherence bet. Pay down the integration story while it's small enough to re-shape.

3. The Answer Isn't Trusting the Agent. It's Never Handing It the Keys. #

The most practical thing I read this week is an engineer's account of running coding agents without giving them any credentials at all. The setup is familiar — several agents at once on disposable VMs, the laptop keeping the SSH keys, the password vault and the CLIs logged in as a real human — and it holds until an agent needs something only the laptop can do. Both standard fixes are bad: copy the token onto the VM and every process there can read it for the token's life, or forward your SSH agent with ssh -A and the VM can use every key you own, without asking, with no record of what it signed.

His tool, lend, proxies the SSH agent socket and adds command execution over it. The agent asks to run gh; the command executes on the laptop under the real login; stdout, stderr and the exit code stream back; the token never moves. The policy is default-deny, per executable and optionally per subcommand, with biometric approval in the dialog and leases that expire when the screen locks. Better still, a hook routes the agent's own prompts — run this, edit that, fetch this URL — into the same dialog, so approvals find you instead of a terminal you stopped watching.

The best section lists what it does not protect against. An approved command runs as your user and isn't sandboxed, so approving a shell — or any interpreter reading from stdin — approves nearly anything. The stated reason comes from the VM and can be faked. Anyone who reaches the forwarded socket can make requests. That honesty is the right shape for the category: the win is narrowing what an agent can ask for and recording what it asked. That is not containment, and anyone selling containment at this layer is selling.

Why it matters:

  • For ICs: Stop running agents as yourself. Disposable compute plus a brokered, approved call back to your credentials is a weekend of setup and kills your worst exposure.
  • For leaders: Agent access should be its own identity with its own audit trail. If agents act as named humans, your incident timeline is already unreadable.
  • For founders: Credential brokering, approval routing and agent identity are unbuilt as products. Demand is arriving faster than the tooling.

The Verdict: Real or Hype? #

Autonomous attack tooling → Real. Named victims, counted records and a three-hour clock in a federal advisory is operations, not a conference demo. AI-driven decoherence as the dominant failure mode → Real but unevenly distributed. Teams that kept architectural ownership aren't feeling it; teams that handed design to the agent are paying already. Per-request credential brokering for agents → Real but early. The pattern is right, the implementations are weeks old, and nothing here sandboxes anything.

── more in #ai-safety 4 stories · sorted by recency
── more on @fbi 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/the-productivity-gai…] indexed:0 read:6min 2026-11-17 · —