AI Brief: Super Intelligence Force, NYC Whistleblowers, and the Shrinking Cyberattack Window

The week's AI news moved on multiple fronts at once: federal governance reshaped, insider safety warnings aired in public, and a new Microsoft report putting hard numbers on how AI is accelerating cyber threats. Alongside those headliners, Google quietly tightened its access tiers while Apple drew a new privacy line for the agentic era.

Trump Establishes a 'Super Intelligence Force' for Federal AI Coordination

The White House has created a new interagency body it is calling the "Super Intelligence Force," tasked with coordinating AI policy across federal departments. The move also accompanies a broader rhetorical shift: official administration communications are replacing the term "artificial intelligence" with "super intelligence" — a framing that signals the administration wants to position the US at the frontier rather than in a regulatory crouch.

The task force draws together representatives from defense, commerce, and national security agencies and is expected to set procurement guidelines, advise on export controls, and push a unified federal stance into international negotiations. Critics have already noted that consolidating AI governance under a single White House-aligned body could sideline independent agency rulemaking, while supporters argue that fragmented departmental approaches have left the US without a coherent posture as rivals accelerate their own programs.

The naming choice is itself a policy statement. "Super intelligence" carries a specific meaning in the technical AI safety literature — it refers to systems that surpass human cognitive ability across all domains — and adopting it as routine government vocabulary suggests the administration sees the race to that threshold as the defining competition of the decade.

Former Anthropic and OpenAI Engineers Warn NYC Council: 'We Cannot Control These Models'

A group of current and former engineers from Anthropic and OpenAI testified before the New York City Council this week, offering some of the most direct public statements yet from industry insiders about the risks of frontier AI development. Several witnesses described internal cultures they characterized as prioritizing capability gains over safety margins, and at least two made explicit reference to extinction-level or civilizational-scale risks.

The core technical claim made during testimony was stark: the companies building the most powerful models cannot reliably predict or constrain model behavior at deployment scale. One witness described alignment research as running years behind capability research, leaving a gap that grows with each new generation of models.

The hearing was convened by the council's technology committee and did not result in binding legislation, but it represents a shift in venue for safety debates that have largely played out in academic papers, conference panels, and internal memos. Bringing these arguments into a municipal legislative chamber — with the record and visibility that entails — marks a new chapter in how AI risk is discussed in public institutions. The testimony drew national press coverage and is likely to inform legislative proposals at both state and federal level in the months ahead.

Microsoft: AI Is Compressing Cyberattack Timelines from Days to Minutes

Microsoft's annual Digital Defense Report, released this week, contains a finding that security teams should treat as urgent: AI-powered threat actors are now executing full attack lifecycles — reconnaissance, initial access, lateral movement, and exfiltration — in minutes rather than the days or weeks that defenders have historically used to detect and respond.

The report documents a tripling of AI-driven phishing incidents, which now account for 23 percent of all phishing attempts tracked by Microsoft's threat intelligence network. These attacks are distinguished by their personalization: AI enables attackers to craft messages that reference accurate contextual details about targets, dramatically improving click-through rates compared with generic lures.

The compression of attack timelines has direct implications for security operations. Detection-and-response workflows designed around hours-long windows are now structurally inadequate. The report calls for AI-assisted defense operating at the same tempo as the attacks themselves — essentially arguing that the security industry needs to match the automation level of the adversaries it faces. For enterprise security teams, the practical takeaway is that mean-time-to-detect targets need to be measured in seconds, not hours, and that human-in-the-loop response models may be too slow for the fastest classes of attack.

Google Restricts Gemini Model Access for Free and AI Plus Subscribers

Starting October 9, Google is narrowing which Gemini models are available to users on free and AI Plus subscription tiers. Free users will be limited to the Flash-Lite model for most interactions, while AI Plus subscribers will see their access to more capable models reduced compared with current entitlements.

The change reflects a pattern emerging across the major AI providers: as inference costs for frontier models remain substantial, companies are using tiered access as the primary monetization lever. Making the most capable models available only at higher price points — or only to enterprise customers — allows providers to serve broad user bases on efficient smaller models while capturing margin from power users and businesses willing to pay for performance.

For developers and researchers who rely on more capable Gemini models for experimentation, the change introduces friction that didn't previously exist. Google has not announced any offsetting improvements to Flash-Lite's capabilities on the announced timeline, meaning some use cases will require a paid upgrade rather than a workaround. The broader industry implication is that the era of frontier-model access as a free or near-free acquisition tool may be ending.

Apple Requires Explicit Authorization Before AI Agents Can Access macOS Full Disk

Apple has updated macOS to require explicit user pre-authorization before any agentic AI application can access Full Disk Access — a system permission that grants broad read rights to documents, mail, calendar data, and other sensitive stores. The change takes effect across the current macOS release and applies to both first-party and third-party AI agents.

The timing is deliberate. As AI agents capable of browsing files, drafting emails, and executing multi-step tasks on a user's behalf move from novelty to mainstream, the attack surface they represent grows substantially. An agent with Full Disk Access and a compromised or misconfigured permission model is effectively a privileged insider threat. Apple's pre-authorization requirement forces a conscious user decision rather than allowing agents to accumulate permissions through background installation flows.

From a developer perspective, the change adds an onboarding step that some users will find confusing — particularly for agents marketed on their ability to work across the filesystem. But the privacy community has broadly welcomed it as an OS-level guardrail that doesn't require users to understand the technical implications of the permissions they're granting. It also sets a precedent: as agents become more capable, expect similar controls at the network, calendar, and contacts layers.