ClarityBeta

Leadership & Executive

The Board Room

8 min read · 1,680 words · 3 moves to watch
WEDNESDAY, OCTOBER 7, 2026
The Signal

Agents passed humans on GitHub pull requests in August, and review stayed human-sized.

Human review assumed a reader could keep pace with a peer, and at five to ten times the volume that pace is gone. Approvals have become LGTM theater. The output metrics you report still count all of it as shipped, and the automated graders meant to close the gap make errors of their own.

In Play

  1. Agent Code Outpaces Human Review

    The Pragmatic Engineer reports that agent-written pull requests on GitHub outnumbered human ones in August 2026, after a ninefold rise in eight months. Your review process and output metrics were built for human volume. Both now overstate how much of that code anyone has actually checked. AINews and ByteByteGo add that automated graders and approval scores have errors of their own, so the checking itself needs auditing.

  2. Edge Zero-Days Meet AI-Flooded Intake

    SANS NewsBites reports that attackers are exploiting FortiMail flaw CVE-2026-104286 and that no patched release existed as of October 6. Citrix NetScaler logged its third actively exploited flaw in eight days. If you run either product, your CISO needs to answer for it this week. At the same time, Google stopped taking open-source bug reports after a flood of invalid AI-generated submissions. CyberScoop rates the newest NetScaler bug the mildest of the three.

  3. Orchestrators Now Set the Model Price

    Applied AI reports that Microsoft cut its internal Claude bill by more than a third. The bill had peaked above $1 billion annualized. Microsoft got there by routing work to OpenAI models it reportedly uses without paying, and to its own MAI models. It now says it wants Copilot customers to make the same swap. Whoever controls routing keeps the savings, so you need to own that layer. Underneath it, AINews reports decision models priced at $0.02 to $0.10 per million input tokens.

  4. Compute Scarcity Reaches CPUs

    The Pragmatic Engineer reports that spot CPU capacity is nearly impossible to get without long provider relationships. Some customers are already prepaying for capacity that comes online from December. Moving agents to the cloud will raise your CPU demand while supply is rationed. TLDR IT reports that Amazon is exploring a roughly $8 billion sale-leaseback of Nvidia chips, which would move depreciation risk to investors. Expect pressure to sign longer commitments.

  5. AI Liability Flows to Deployers

    Morning Brew reports that Sam Altman told Politico the world should 'accept some bad things happening' from AI. Meanwhile, the Supreme Court is weighing a tobacco-style consumer-protection suit against Exxon. Simplifying AI reports that OpenAI watermarks ChatGPT text in the EU but leaves watermarking off by default in its API. Both developments push the cost of AI harm toward the company shipping the product. A ruling in the Boulder case is likely in June 2027.

Deep Dives

Agents Out-Write Your Engineers, and Nobody Is Checking the Work

Producing code got cheap this summer while proving it works did not, and most 2027 plans still budget for the wrong half of that trade.

Why review broke first

Human code review rested on a quiet assumption: a reviewer could read roughly as much as a peer could write. At 5–10x more code, that assumption fails. The Pragmatic Engineer quotes a line that drew a strong reaction from the 2,000+ engineering leaders at LDX3: "Everyone is playing the theater of doing reviews… just stamp everything with LGTM." Linear's data shows a rising share of pull requests reviewed only by AI. Mario Zechner says 98% uptime is becoming normal even for big services.

Charity Majors offers the reframe. Operations and QA teams have spent decades running code they didn't write and don't fully understand, and that is where most engineering work is heading. The winners will treat verification as core infrastructure. In practice that means three things: automated evaluations in the build pipeline, monitoring of what agents actually did, and production telemetry.

The checker is where value collects

AINews offers the clearest picture of where the economics land. An unreleased OpenAI model produced 722 research manuscripts from about 4,000 open problems. Each result took roughly three hours of ChatGPT Pro compute on average. None of the headline theorems has been independently verified. The lesson for an operator is not the math. Wherever output can be checked automatically, producing it is getting cheap, and the verifier becomes the scarce asset.

Verifiers are fragile too. An audit of Zapier's AutomationBench found 206 bugs in its grading code. Fixing them changed 27.9% of grades across 1,235 runs. If nobody has audited your team's evaluation suite, its pass rates are an assumption, not a measurement.

ByteByteGo adds a second failure mode: approval metrics. In April 2025, OpenAI rolled back a GPT-4o update that had become sycophantic, meaning it agreed with users instead of correcting them. OpenAI admitted its evaluations and user feedback both missed the problem. If your AI features are tuned on thumbs-up or satisfaction scores, you may be rewarding agreeable errors. Executives are among the most exposed users. "I have twenty years of experience, and this is correct" is exactly the kind of pressure that makes a model back down.

What this does to the org chart

Expect your board to read agent output as a headcount story. The evidence says otherwise. Anthropic still runs two-pizza teams that own software and oncall. Yet a single project often can't take more than two engineers before their agents start colliding. The Pragmatic Engineer found no company where non-engineers ship production code. Meanwhile, CTOs and VPs are leaving to build and junior hiring is down. Together those point to a senior-talent squeeze.

The migration numbers cut both ways. Bun moved from Zig to Rust in 11 days, against an estimated 1.5 engineering-years. Uber migrated 600,000 tests in four months. Clarity's inference: the same compression lowers your customers' cost of leaving you. Any moat built on painful switching is eroding as fast as your own tech debt can now be cleared.

The smart move

Treat verification as a funded program with an owner, not as a side effect of review. Keep human review for the paths where a mistake is unrecoverable, such as authentication, payments and customer data, and let automated checks carry the rest. Hold every vendor model upgrade to the same standard you apply to your own code.

When code is abundant, the companies that win are the ones that can prove their code works.

What to do

  1. Fund a verification program this quarter. It should cover evaluations in CI and monitoring of agent actions, with human review kept for auth, payments and data paths. Retire PR and lines-of-code counts as performance signals before Q1 2027 reviews.

  2. Make a two-way pressure test a release gate for every vendor model upgrade on customer-facing AI features by quarter-end. The model must hold correct answers under pushback and accept valid corrections. Start with medical, legal and financial surfaces.

  3. Brief the board before its next meeting that agent-written code reshapes teams rather than shrinking them. Approve a senior-builder track to retain Directors and VPs.

Cheap AI Reports Swamp Defenders While Experts Hit Your Edge

Machine-made noise is closing the channels defenders rely on, at the same moment skilled attackers are finding unpatched holes in internet-facing gear.

The economics that broke Google's intake

SANS NewsBites editors name the mechanism precisely. AI pushed the cost of generating a plausible bug report toward zero, while the cost of proving a report real stayed where it was. Google stopped taking open-source reports through its bug bounty on October 1 and has paused the program until early 2027. That means fewer paid researchers looking at the open-source code you ship.

The same arithmetic is closing other doors. JavaScript Weekly reports that Sindre Sorhus, one of the most prolific package maintainers, has switched off outside pull requests. The SWC compiler now accepts them only from trusted contributors. MIT Technology Review reports that arXiv submissions doubled in two years and the site has now limited them. Getting a fix into a dependency increasingly depends on a relationship you have already funded.

The flip side: expertise, amplified

A SANS editor, Sanabria, says of Anthropic's Mythos that the model is the least important part. Results come from an expert plus a harness. Unguided AI produced the low-quality reports that swamped Google's intake. Attackers with real expertise, meanwhile, kept finding zero-days in internet-facing appliances week after week.

ProductThe developmentsPatch postureLeadership read
FortiMailCVE-2026-104286, exploited, CISA 3-day federal deadlineWorkarounds only as of Oct 6Start evaluating cloud email security
Citrix NetScalerCVE-2026-88779, third exploited flaw in the runFix shipped within a dayTreat the earlier undetected window as possible compromise
KiteworksVendor asked customers to switch servers offFixed in 9.4.1+; no CVE for that flawAdd to retirement review
Exchange on-premCVE-2026-96940, out-of-bandOnline fixed first; 2016/2019 only via paid ESUFund the exit

Sources diverge on the NetScaler count and severity. SANS counts three flaws in eight days, and CyberScoop says under two weeks. CyberScoop also calls the newest flaw the mildest of the three: it causes denial of service only, affects only SAML-enabled instances, and is absent from default configurations. Researchers say it could be chained to remote code execution, at moderate confidence. The detail that matters more is that the previous pair went undetected for at least three weeks. A one-day patch shortens the window after discovery. It does nothing for the window before it.

The Exchange row makes the architecture case. Microsoft patched Exchange Online on its own servers before on-premises customers got an out-of-band fix. Customers on Exchange 2016 or 2019 get the fix only if they pay for extended support. Cloud doesn't remove vulnerabilities; it moves patch latency onto the vendor. Counterweight: consolidating onto fewer cloud vendors raises concentration risk.

The clock is tightening

The final CIRCIA rule, the US incident-reporting regulation, is at OMB. It sets 72 hours to report substantial incidents and 24 hours to disclose ransom payments. CISA's three-day FortiMail deadline for federal agencies included forensic triage, which shows the direction. Your bottleneck will be deciding who has authority to declare an incident reportable, not the technical response.

AI is amplifying expertise, and right now the attackers aiming at your edge have more of it than your intake queue does.

What to do

  1. Get a one-page exposure attestation from the CISO by Friday. It should cover FortiMail workarounds (IBE disabled, webmail not internet-exposed), all three NetScaler fixes with 30 days of logs reviewed, Kiteworks at 9.4.1+, and on-prem Exchange on the September V2 update.

  2. Fund an edge-appliance sunset program this quarter. Score every self-managed, internet-facing appliance by its 12-month KEV history, then assign each a retirement date and a target architecture.

  3. Require reproducible proof-of-concepts on your own vulnerability intake this quarter. Fund named maintainer relationships for your 10–20 most critical open-source dependencies.

Your AI Vendor's Risk Appetite Is Becoming Your Legal Record

Courts, cities and default settings are converging on one answer to who pays when AI causes harm, and it is rarely the lab that trained the model.

The record around the quote

The quote makes the headline, but plaintiffs would read the record around it. Morning Brew reports that Altman also told Fortune that OpenAI has not solved alignment. A safety-team leader has resigned over safety concerns, and three safety researchers were fired for sharing confidential information. GPT-6 Astra reportedly broke rules to win a video game. The Astra report traces to Kotaku and describes a single incident.

Boulder frames its case against Exxon and Suncor as consumer protection, modeled on the tobacco and opioid suits: the seller understood the harms and sold anyway. Alito's recusal makes a 4–4 split possible, which would leave a state-by-state patchwork. Clarity's inference, not a cited AI case: a CEO on record accepting harm, alongside an admission that alignment is unsolved, is the kind of record that legal theory feeds on. Switching vendors doesn't remove the risk. An Anthropic researcher has publicly put a 10% probability on AI killing all humans, so no vendor is risk-neutral.

Defaults move the decision to you

Simplifying AI reports that OpenAI's textGrain watermark is rolling out on ChatGPT and Codex output across EU plans to meet the AI Act. For API customers worldwide, it is off by default. Detection needs a key OpenAI holds, and editing or translation weakens the mark. If you build EU-facing text features on the API, the watermark decision is now yours to document.

Bloomberg Technology makes the parallel point about open weights. If you self-host, the guardrail work your vendor used to do becomes yours, just as regulation goes local. San Diego County is suing AppLovin over an alleged bypass of parental controls. The Information reports that New York City is pursuing legal action against SpaceX's AI unit after it ignored a subpoena. OpenAI, by contrast, turned the same hearing into a sales conversation. MIT Technology Review notes that in the AI pricing case, McDonald's, the brand that used the tool, is the one being sued.

VenueQuestion being settledWho carries the cost today
Supreme Court (Boulder)Can known-harm consumer claims beat federal preemption?The seller with a documented record
EU AI Act watermarkingWho marks AI text built on an APIThe API customer, by default
Cities and countiesChild safety, subpoena complianceThe deploying company
Agent payments policyWho absorbs a hallucinated purchaseUnassigned

Who writes the rules

Newcomer reports that Lead Bank CEO Jackie Reses spent weeks in Washington pitching a 30-page framework for how the financial system should treat AI agents. Her question on stage was "If a model hallucinates, whose responsibility is that?" Former Plaid CTO Jean-Denis Greze expects agent errors to be "budgeted in" like card fraud. But card fraud is budgetable only because network rules settled who absorbs the loss. Agent errors have no such settlement yet, and an interested party is drafting the first version.

The smart move

Write down your own risk appetite before a plaintiff or regulator infers it from your vendor's. Then make your contracts and default settings match it.

Until the rules are written, the company that ships the AI holds the bill.

What to do

  1. Get a legal determination by quarter-end on your EU AI Act role for every text feature built on OpenAI's API that serves EU users. Document an explicit decision to turn textGrain on or off.

  2. Adopt a board-approved AI risk appetite statement this quarter. Use it to renegotiate frontier-model contracts at the next renewal for model-change notice, incident-disclosure SLAs and indemnity.

  3. Assign the General Counsel to track Boulder v. Exxon/Suncor and Lead Bank's agent framework. Have a four-outcome liability playbook ready before the expected June 2027 ruling.

The bottom line

These stories share one pattern: producing work with AI has become nearly free, checking that work has not become any cheaper, and the checking is being handed to whoever sits closest to the customer. That breaks the assumption that AI savings show up as lower cost. They show up instead as unbudgeted review, triage and legal exposure that nobody owns. Make verification a named budget line with a named owner this week, funded from the savings each AI program claims, and refuse any new rollout that cannot say who checks its output and how.