6 min read · 12 sources

Anthropic has formally filed its registration for a US IPO, offering the market its first comprehensive look at the capital reality behind frontier foundation models. The headline figure is eye-watering: an $80.7 billion net loss for the first half of the year.

Most of that paper loss stems from non-operating convertible-bond accounting charges tied to skyrocketing private valuations. Operationally, the picture is tighter: Anthropic’s operating deficit narrowed to $693 million as enterprise API revenue scaled up. Still, the prospectus makes no secret of the severe risks ahead, underlining climbing compute expenses and deep concentration risk around a tight circle of chip vendors and cloud hosting providers.

AWS’s 1:1 NAT lets overlapping 10.0.0.0/16 networks talk without renumbering or dedicated appliances.

Gheist exploits stolen GitHub tokens to exfiltrate private repository secrets

Source: wiz.io ↗

A newly documented automated campaign named Gheist is burning through compromised GitHub personal access tokens to extract pipeline credentials at scale. Active since mid-September 2026, the tool targets enterprise and private environments by pushing branches named ci-sync-[random] and committing an unsigned .github/workflows/sync-check.yml workflow file.

The exploit relies on a classic runner blind spot: masking workarounds. To bypass GitHub Actions’ automatic masking of known secrets, the injected workflow base32-encodes raw secret values, dumps the encoded strings straight into the execution logs, and retrieves them via the GitHub API.

While GitHub’s default fork-PR approvals successfully neutered Gheist runs on public repositories, Wiz observed successful executions against private corporate repos where those controls were missing. The campaign actively pillaged environment-scoped deployment secrets to pivot into target AWS and Azure footprints. SREs and AppSec teams should scan access logs for the gheist API user-agent string, audit existing token scopes, and ensure branch-protection rules enforce mandatory reviews on all workflow file changes.

Microsoft guts standalone Azure Communication Services

Source: bloggeek.me ↗

Microsoft is pulling the plug on generic customer-facing telecom infrastructure, confirming the retirement of major Azure Communication Services components by September 30, 2028. As of October 23, 2026, Microsoft has halted all new onboarding for SMS, Email, WhatsApp, and standalone video calling.

What remains will be shoehorned directly into the Teams ecosystem. Surviving APIs are migrating to Teams-centric patterns - such as Teams Phone Extensibility - delivered through breaking 4.x SDK upgrades. If your engineering team built customer support chat, transactional SMS notifications, or in-app telemedicine video on top of ACS, you are officially on a two-year migration clock to alternative CPaaS vendors like Twilio or Vonage.

AWS launches managed 1:1 NAT on Transit Gateway and Cloud WAN

Source: aws.amazon.com ↗

Connecting two VPCs or on-premises data centers that share the exact same 10.0.0.0/16 address space has historically been an operational headache, requiring fleets of self-managed EC2 NAT instances, complex iptables rules, or outright IP renumbering. AWS has finally resolved this by rolling out native 1:1 NAT on AWS Transit Gateway and AWS Cloud WAN.

The new capability integrates directly with Regional Private NAT Gateways across all commercial regions. The managed fabric automatically handles bidirectional translation for overlapping private CIDR blocks across VPC attachments, Direct Connect, Site-to-Site VPNs, and software-defined WAN appliances. Network engineers can provision rules via CloudFormation, the AWS CLI, or the console without configuring custom return-route hacks, with full metric instrumentation exposed in CloudWatch and EventBridge.

Consumer agents establish unmanaged OAuth footholds in enterprise SaaS

Source: reco.ai ↗

A new threat vector is emerging from the executive suite: unmonitored consumer AI agents linking to core SaaS infrastructure. Reco discovered widespread enterprise Google Workspace, Microsoft 365, and Slack exposure caused by employees granting high-privilege OAuth scopes to consumer assistants like Instinct, Meta’s Muse, and Grok.

The architectural risk is structural. Instinct - a consumer personal assistant that reached a $10 billion valuation within months of its launch - uses interconnected “Trusted Person Networks” that dynamically sync calendar data, address books, and internal file drives between users. When an executive hooks the agent up to their corporate inbox, they effectively open a bidirectional, non-human backdoor that ingests sensitive corporate data into an unvetted third-party multi-tenant graph. Security teams need to immediately pull third-party application audit logs and clamp down on API client ID whitelisting.

Operations and automation: Monday.com, Cisco Wi-Fi 7, and Outlook

Source: engineering.monday.com ↗

At monday.com, production engineers faced a scaling wall: internal automated PR generation pushed on-call ticket volume to 10x human engineering capacity. Rather than hiring a small army of operators, monday.com built an internal LLM agent platform with strict production guardrails. The agent loop ingests ticket context and leverages constrained APIs to read, write, and validate Terraform, Helm, and GitOps configurations. Over one quarter, the system autonomously resolved roughly 50 percent of the team’s 1,688 monthly infrastructure tickets without escalation.

Cisco is shifting its enterprise wireless focus toward continuous agent enforcement with its new CW9176ISEC Wi-Fi 7 access point. The hardware introduces post-quantum cryptography aligned to CNSA 2.0 standards, alongside refreshed Catalyst 9800 Series controllers running Auto MACsec to secure traffic between access points and switches. The design anticipates agent-heavy enterprise networks where devices continuously traverse physical access points while initiating persistent API sessions.

Meanwhile, Microsoft is closing a long-standing parity gap by adding a unified “All Accounts” inbox to the new Outlook for Windows. Rolling out globally through late October 2026, the client aggregates messages from multiple connected IMAP, Exchange, and personal accounts into a single view without merging underlying mail stores. Centralized administrative policy controls will roll out later this month.

Industry perspectives: Metric traps, defense sandboxes, and cognitive mercantilism

Source: cio.com ↗

Chasing aggregate service availability numbers remains one of the easiest ways for engineering leadership to mislead themselves. An analysis on vanity IT operational metrics points out that hitting 99.9% uptime or strict ticket-resolution SLAs regularly masks catastrophic failure modes. Teams optimize for closing alerts quickly using brittle workarounds while chronic edge-case corruption festers unnoticed. The fix requires orienting dashboards around business disruption and repeat incidents rather than aggregate uptime targets.

In federal tech, the U.S. Army Office of the CIO launched Project Gatekeeper to overhaul defensive software procurement. The automated DevSecOps sandbox tests containerized commercial software in minutes, bypassing legacy Authority to Operate (ATO) certification processes that traditionally dragged on for six to eight months. Commercial software vendors can now push images into the testing pipeline and generate automated compliance dossiers for military network deployment.

In HR tech, enterprise provider Dayforce agreed to acquire conversational hiring startup Ribbon AI, integrating voice and video screening agents directly into its platform to run dynamic, transcript-scored initial candidate interviews.

Finally, an essay by Venkatesh Rao critiques current AI doomerism through the lens of economic history, dubbing the prevailing cultural panic cognitive mercantilism. Rao argues that contemporary critics view human intelligence as a fixed, finite pie that automation must inevitably pillage - replaying 17th-century zero-sum mercantile fallacies. In contrast, widespread machine intelligence is poised to unlock an industrious revolution of abundant, net-new technical demand.