Friday’s 5 PM technology recap follows five developments with immediate lessons for resilient systems, safer hiring, credible AI oversight, practical model selection, and device purchasing. A millisecond software defect disrupted UK air traffic, a multinational advisory detailed the WaterPlum recruitment campaign, more than 100 specialists set conditions for independent AI evaluation, enterprise teams described why last year’s models can still be enough, and a teardown found Qualcomm hardware inside one new iPhone configuration.[1][3][5][7][9]
Reporting cutoff: September 18, 2026, 5:09 p.m. EDT (21:09 UTC). Every selected development was published after the prior 5 PM cutoff. NATS calls its finding preliminary; WaterPlum attribution and victim figures are joint-agency assessments; the AI-evaluator letter is a proposal, not an adopted rule; Dreamforce interviews are not a market-wide survey; and the iPhone teardown covers an examined unit while broader deployment details remain under study.
Listen to the broadcast-ready recap
This KittenTTS Bella narration was synthesized locally from the same checked facts as the article.
Tonight’s five-story rundown
1. A millisecond defect shows why resilient software needs more than uptime
NATS published a preliminary report saying the September 8 disruption began when a manual aircraft-identification-code request was paused by a higher-priority activity. When the request resumed, a software defect caused corrupted output that affected later flight-data updates; NATS says the triggering sequence occurred within a millisecond.[1]
Reduced information led controllers to restrict traffic while the National Airspace System was restarted and flight data reloaded. NATS says restrictions lasted about six hours and the passenger backlog took more than two days to clear. Reuters reported about 2,000 cancellations.[1][2]
Practical action: teams running critical systems should test interrupted and resumed transactions, priority inversions, corrupted-state recovery, and whole-system restart procedures—not merely normal requests. A fail-safe restriction can protect people while still causing major operational harm, so recovery capacity matters alongside fault detection.
2. WaterPlum turns job interviews into an attack path
A September 18 advisory from US, Japanese, Australian, and German agencies assesses that North Korean WaterPlum actors posed as recruiters or employers in AI, blockchain, cryptocurrency, and related fields, then pushed malicious coding assignments or troubleshooting files to developers. The advisory describes malware including BeaverTail, InvisibleFerret, OtterCookie, OtterCandy, and StoatWaffle.[3]
The agencies assess that activity from December 2025 through July 2026 infected at least 30,000 devices in more than 100 countries, affected funds or account credentials tied to more than 7,000 cryptocurrency wallets, and transferred 1.7 billion Japanese yen—about $10.71 million—to North Korea.[3] The Record independently confirmed the advisory’s publication and reporting context, but it did not reconstruct those totals from individual victims.[4]
Practical action: never run interview code on a production workstation or a device holding wallets, credentials, or customer data. Open unfamiliar Visual Studio Code projects in Restricted Mode, inspect task files with a separate editor, isolate tests in a disposable environment, and give contractors only the access required for the assignment.[3]
3. Independent AI evaluation needs independence in practice
The AI Evaluator Forum published a letter signed by more than 100 experts calling for embedded third-party evaluators to retain scientific objectivity, editorial control, transparent methods, access to relevant systems and staff, and protection from company interference or retaliation.[5] CNBC reported that signatories include Geoffrey Hinton and people affiliated with universities and evaluation organizations including Stanford and METR.[6]
The proposal also says evaluators should disclose and mitigate conflicts, avoid compensation contingent on findings, communicate with boards and oversight bodies, and be able to release findings subject to bounded handling of genuinely sensitive material. The letter says embedded evaluation should complement—not replace—broader transparency and independent research.[5]
Practical action: buyers should ask who selected and paid an evaluator, what systems were accessible, whether negative findings could be published, which conflicts were disclosed, and whether the testing covered deployment and incident practices rather than only a polished model demo.
4. The useful AI model may be the cheaper one you already understand
Salesforce customers and partners interviewed by CNBC at Dreamforce said older and cheaper models can handle many everyday sales and customer-service tasks. Docusign CEO Allan Thygesen described model routing: sending each request to the most cost-effective system rather than defaulting every task to the newest frontier model.[7]
Docusign’s engineering account says its AI pipeline processes more than one million documents daily, selects model tiers according to the extraction task, validates output before it reaches a system of record, and measures per-document inference costs in fractions of a cent.[8] Those are company-reported production figures, not an independent audit of every workload.
Practical action: benchmark complete business outcomes: accuracy, exception rate, latency, data handling, reviewer time, and cost per accepted task. Reserve expensive models for work that measurably benefits from them, and keep smaller or specialized models where they meet the requirement.
5. One iPhone 18 Pro Max teardown finds Qualcomm still inside
TechInsights’ preliminary teardown of an iPhone 18 Pro Max model A3473 identified Apple’s A20 Pro processor and N1 wireless module alongside Qualcomm’s Snapdragon X80 modem and a newly identified Apple mmWave antenna module.[9]
MacRumors reported that the examined configuration is the US iPhone 18 Pro Max, while the smaller Pro, iPhone Duo, and non-US Pro Max configurations use Apple’s C2 modem. Apple has not publicly explained the split, and TechInsights says more package-level and device analysis is needed before drawing conclusions about Apple’s broader modem strategy.[9][10]
Practical action: buyers who care about weak-signal behavior, battery use, roaming, or mmWave should wait for controlled testing on their exact regional model and carrier. Confirm the model number before relying on a teardown from another market.
What to do next
- Critical-system teams: add interruption, resume, priority, state-corruption, and recovery tests to the reliability plan.
- Developers and hiring teams: isolate interview code, keep editor trust disabled by default, verify identity, and minimize contractor access.
- AI buyers: ask for evaluator independence and benchmark task outcomes before paying for the newest model.
- Phone buyers: verify the regional model and wait for direct radio and battery testing.
Need help securing, fixing, or planning your technology? Schedule The IT Guys
More from The IT Guys
Frequently asked questions
Did the NATS software defect put aircraft in danger?
NATS says safety was not compromised and that traffic restrictions were imposed because controllers had reduced information. That is the operator’s preliminary account; the independent review remains pending.[1]
Is every coding interview repository dangerous?
No. The lesson is to verify identity and isolate unfamiliar code because WaterPlum used normal-looking recruitment workflows and developer platforms as delivery paths.[3]
Does the AI evaluator letter create a new rule?
No. It proposes minimum conditions for credible embedded evaluation. Adoption, enforcement, and the labs’ detailed responses remain unresolved.[5]
Are older AI models always cheaper overall?
No. Inference can cost less, but total cost also includes accuracy failures, retries, human review, integration, privacy controls, and incident risk.
Does the Qualcomm modem make the US iPhone better?
The teardown identifies the component; it does not establish a performance winner. Controlled regional testing is still needed.[10]
Sources checked for this edition
- NATS — NATS publishes preliminary report on technical incident of 8 September
- Reuters via Devdiscourse — UK air traffic control outage caused by software defect, says operator
- FBI, DC3, and partner agencies — North Korean “WaterPlum” advisory
- The Record — North Korean hackers infect thousands of devices across 100 countries as part of “WaterPlum” campaign
- AI Evaluator Forum — Minimum Conditions for Embedding Evaluators
- CNBC — Anthropic and OpenAI need truly independent safety evaluators, experts say in public letter
- CNBC — AI safety debate meets reality at Dreamforce as business leaders say last year’s models are enough
- Docusign — Architecting for Millions: How Docusign Processes Agreements at Scale
- TechInsights — Apple iPhone 18 Pro Max Teardown: A20 Pro, X80 and Apple N1
- MacRumors — iPhone 18 Pro Max in U.S. Has Qualcomm’s Snapdragon X80 Modem
Source state checked through September 18, 2026, 5:09 p.m. EDT / 21:09 UTC. Primary and independent claims are labeled separately; preliminary findings, agency assessments, company figures, proposals, and unknowns remain qualified.