Anthropic's Compute Costs Rise, OpenAI Automates Linux Exploit
This digest covers Anthropic's increasing compute expenses, OpenAI's automated Linux root exploit generation, and new Kubernetes rightsizing tools from
The latest news highlights the escalating costs and increasing sophistication in the AI landscape. While AI models demonstrate powerful capabilities in vulnerability discovery, the underlying compute infrastructure continues to be a major expense. Concurrently, new security threats like LLMJacking are emerging, underscoring the critical need for vigilance.
TL;DR
- OpenAI Pipeline Generates 948 Linux Root Escalation Exploits in 30 Hours for $1.2M — OpenAI's AI pipeline, using an upcoming cyber-focused model, generated 948 Linux local privilege escalation exploits within 30 hours for $1.2 million in API costs.
- Google Cloud Reports Rise of LLMJacking: Threat Actors Monetize Compromised LLM Resources — Google Cloud's Mandiant reports a rise in 'LLMJacking,' where threat actors hijack enterprise access to LLMs for fraudulent resource consumption and monetization.
- CISA Adds Exploited Apple Flaw CVE-2024-27834 to KEV Catalog Amid Active Exploitation — CISA added a medium-severity Apple vulnerability, CVE-2024-27834, to its KEV catalog due to active exploitation by APT29 to deliver info-stealing malware.
- Kremlin-Linked Hackers Exploited Patched iOS and Android Flaws in Watering Hole Attacks — Hackers, likely linked to the Kremlin, exploited previously patched iOS and Android vulnerabilities via watering hole attacks to infect diplomatic targets.
- APT29 Uses Hijacked Web Analytics Script for iOS/Android Exploits in Poisoned Source Campaign — APT29 is using a supply chain compromise of the Plerdy SEO Tracker to inject malicious JavaScript and deliver iOS and Android exploits in a campaign named Poisoned Source.
- Anthropic's Compute Costs Rise to 58% of Revenue, Impacting Margins Ahead of IPO — Anthropic's compute costs increased to 58% of its revenue in the most recent quarter, up from 48%, negatively impacting its operating margins.
- Meta Acquires AI Coding Startup Conductor to Enhance Its AI Capabilities — Meta has acquired Conductor, an AI coding startup specializing in deploying multiple AI agents into parallel workspaces for software projects.
- Groundcover Releases Kubernetes Rightsizing Beta from Wand Acquisition — Groundcover has launched a beta for its Autonomous Infrastructure feature, which automatically rightsizes Kubernetes CPU and memory requests based on production behavior.
- Discovered Materials Launches $1M Challenge for AI to Solve Materials Science Problems — Discovered Materials, a Y Combinator startup, is offering $1 million in prizes for AI models to solve five critical, unsolved problems in materials science.
- Anthropic Develops Method to Authenticate AI Model Responses with High Compute Cost — Anthropic has developed a method to cryptographically prove which AI model generated a response, but it comes with a significant compute cost.
OpenAI Pipeline Generates 948 Linux Root Escalation Exploits in 30 Hours for $1.2M
Top story · AI Models · The Hacker News · 2026-11-05
OpenAI has detailed a new AI pipeline that generated 948 local privilege escalation (LPE) exploits for Linux systems in a 30-hour run, costing $1,213,430 in API credits. The pipeline, powered by an upcoming cyber-focused AI model, identified and analyzed 2,308 candidate Linux kernel findings and dynamically reproduced 1,724 vulnerabilities.
This system expands upon previous work by OpenAI's GPT-5.5-Cyber model. The five-stage pipeline includes vulnerability discovery, dynamic validation in a sandboxed environment, exploitability assessment to create proofs-of-concept, patch generation, and human review. The process also uncovered at least 10 novel Linux exploitation techniques, including new persistent elastic objects, page grooming methods, data-only privilege escalation techniques, and an arbitrary type confusion method.
Key facts
- 948 local privilege escalation (LPE) exploits generated.
- Total run time: 30 hours.
- Total API cost: $1,213,430, or $1,279 per LPE exploit.
- Identified 10 novel Linux exploitation techniques.
Why it matters: This demonstrates AI's accelerating capability in identifying and exploiting critical vulnerabilities, requiring systems and security engineers to drastically accelerate their patching and security hardening processes to keep pace with AI-driven attack vectors.
AI is now an industrial-scale exploit generator, shifting the cybersecurity arms race dramatically towards automated defense and patching.
🔗 Read more at The Hacker News
Google Cloud Reports Rise of LLMJacking: Threat Actors Monetize Compromised LLM Resources
AI Agents & Tooling · Google · 2026-11-03
Google Threat Intelligence Group (GTIG) and Mandiant Services have observed a new trend called "LLMJacking," where threat actors compromise enterprise access to Large Language Models (LLMs) across various cloud and LLM providers. Incidents in 2026 impacted organizations in North America, South America, and Asia Pacific, spanning financial services, retail, technology, and transportation sectors. This tactic leads to rapid, large spikes in fraudulent cloud charges, with some organizations seeing over $100,000 USD in a single day and cumulative charges reaching $1 million USD.
Unlike cryptojacking, LLMJacking focuses on directly targeting backend model resources by identifying publicly exposed LLM endpoints or obtaining valid LLM API keys. An emerging criminal ecosystem includes "LLM Mixers," illicit routing gateways that proxy, aggregate, and resell unauthorized LLM access. GTIG observed specialized operational clusters coordinating initial access and monetization, demonstrating increasing sophistication in this threat landscape. Mandiant provides detection, hardening, and remediation guidance.
Key facts
- LLMJacking observed across multiple incidents in 2026.
- Fraudulent cloud charges exceeded $100,000 USD in a single day.
- Cumulative charges reached $1 million USD at some organizations.
- Threat actors target publicly exposed LLM endpoints or valid LLM API keys.
- Emergence of 'LLM Mixers' for reselling unauthorized LLM access.
Why it matters: AI and backend engineers must immediately review their LLM deployments for publicly exposed endpoints, strengthen API key management, and implement robust monitoring to detect anomalous LLM resource consumption to prevent significant financial losses and data exfiltration.
LLMJacking is a new, costly threat where compromised LLM access leads to massive fraudulent cloud charges, demanding urgent security attention for LLM deployments.
CISA Adds Exploited Apple Flaw CVE-2024-27834 to KEV Catalog Amid Active Exploitation
Security · The Hacker News · 2026-11-05
The U.S. Cybersecurity and Infrastructure Security Agency (CISA) has included CVE-2024-27834, a medium-severity Apple security flaw, in its Known Exploited Vulnerabilities (KEV) catalog. This action follows reports of active exploitation by threat actors, possibly linked to Russia's Foreign Intelligence Service (APT29), to deploy information-stealing malware.
The vulnerability is an improper authentication issue that could allow an attacker with arbitrary read and write capabilities to bypass Pointer Authentication. Apple addressed this flaw in May 2024 with improved checks in iOS 16.7.8, iPadOS 16.7.8, iOS 17.5, iPadOS 17.5, macOS Sonoma 14.5, tvOS 17.5, watchOS 10.5, and Safari 17.5. The exploit chain involves CVE-2025-43529, a WebKit remote code execution flaw, combined with the PAC bypass to execute JavaScript for browser cookie theft. Federal Civilian Executive Branch agencies are advised to apply patches by November 18, 2026.
Key facts
- CVE-2024-27834 (CVSS score: 5.5) is an improper authentication flaw.
- Apple patched the vulnerability in May 2024 (iOS 16.7.8+, macOS Sonoma 14.5+, Safari 17.5+).
- Exploitation attributed to APT29 (Cozy Bear/Midnight Blizzard) with moderate to high confidence.
- FCEB agencies recommended to patch by November 18, 2026.
Why it matters: Backend and systems engineers must ensure all Apple devices and systems under their purview are updated to patched versions to mitigate the risk of info-stealing malware and protect against sophisticated state-sponsored attacks.
Actively exploited Apple flaws demand immediate patching to prevent advanced persistent threat groups from exfiltrating sensitive data.
🔗 Read more at The Hacker News
Kremlin-Linked Hackers Exploited Patched iOS and Android Flaws in Watering Hole Attacks
Security · Ars Technica · 2026-11-05
Hackers, assessed with "moderate to high confidence" to be part of Russia’s Foreign Intelligence Service (APT29), exploited previously patched iOS and Android vulnerabilities. These attacks targeted visitors of several websites, including one impersonating america250.org, a site celebrating American independence, and theowo.club, which proxied a luxury hotel near UK government institutions.
The attack methodology involved injecting malicious JavaScript into legitimate website traffic or using proxy domains. This JavaScript profiled the visitor's device before delivering platform-specific zero-click exploit chains. The payloads aimed to steal browser cookies, Wi-Fi passwords, and full mobile device data. One of the iOS exploit kits used was DarkSword, which Google researchers previously noted as being used by commercial surveillance vendors.
Key facts
- Attacks attributed to Russia's Foreign Intelligence Service (APT29).
- Exploited previously patched iOS and Android flaws.
- Used watering hole attacks via compromised websites, fake domains, and supply chain compromise.
- Targeted sites included america250.org impersonations and a luxury hotel near UK government buildings.
Why it matters: This confirms that even patched vulnerabilities can be exploited in sophisticated campaigns, stressing the need for continuous monitoring, threat intelligence, and a defense-in-depth strategy, especially for mobile device management and web analytics services.
State-sponsored actors are actively exploiting patched mobile vulnerabilities through elaborate watering hole attacks, emphasizing layered security and prompt patching.
APT29 Uses Hijacked Web Analytics Script for iOS/Android Exploits in Poisoned Source Campaign
Security · The Hacker News · 2026-11-04
The Russia-linked threat actor APT29 (Cozy Bear) is conducting a large-scale watering hole campaign called "Poisoned Source," targeting iOS and Android users. This campaign leverages compromised legitimate websites, fake lookalike domains, and a supply chain compromise of the Ukrainian web analytics service Plerdy SEO Tracker. The malicious code was injected into "a.plerdy[.]com/public/js/click/main.js" between August 27 and September 3, 2026.
The injected JavaScript profiles the visitor's device and browser, then delivers platform-specific N-day exploit chains. The ultimate goal is to steal browser cookies, and in some cases, Wi-Fi passwords or full mobile device data. High-value targets, including military-related websites and diplomatic visitors, are suspected, although it's unclear if the aim is broad compromise or specific targeting. For iOS, the DarkSword exploit kit is used, while Android malware exfiltrates Chrome data and Wi-Fi passwords without persistence.
Key facts
- APT29 is behind the "Poisoned Source" campaign.
- Supply chain compromise of Plerdy SEO Tracker (a.plerdy[.]com/public/js/click/main.js) from August 27 to September 3, 2026.
- Three infection vectors: compromised sites, lookalike domains (e.g., america250[.]holiday), and supply chain.
- Uses DarkSword exploit kit for iOS and malware for Android to steal data.
Why it matters: Backend and systems engineers using third-party web analytics or similar services must scrutinize their supply chain for potential compromises, as these can be a vector for highly targeted and persistent attacks from state-sponsored actors.
Supply chain compromises of seemingly innocuous services like web analytics scripts are a critical vector for sophisticated mobile device exploitation.
🔗 Read more at The Hacker News
Anthropic's Compute Costs Rise to 58% of Revenue, Impacting Margins Ahead of IPO
Industry · TechCrunch · 2026-11-03
Anthropic's latest financial filing, an update to its S-1, reveals that compute costs consumed 58% of its revenue for the quarter ending September 30th, up from 48% in the prior quarter. Despite revenue growing to $17 billion from $11.5 billion, the operating loss also increased due to these rising compute expenses.
The company's compute bill for the quarter reached $10 billion, including $5.0 billion spent on short-term commitments at higher rates, indicating a scramble for resources to meet demand. Anthropic has deals with providers like SpaceX ($1.25 billion/month), Nscale, Akamai, and Volta, often using 'neoclouds' for flexibility. The company expects compute expenses to remain its largest operating cost even as multi-year arrangements become available, and its efforts to build its own compute base, 'Anthropic Direct,' cannot keep pace with demand.
Key facts
- Compute costs rose to 58% of revenue in the latest quarter (Q3).
- Revenue for the quarter reached $17 billion, up from $11.5 billion.
- Quarterly compute bill was $10 billion, including $5.0 billion on short-term commitments.
- Signed deals with SpaceX ($1.25 billion/month), Nscale, Akamai, and Volta.
Why it matters: This highlights the immense and rising operational expenses for AI model development and deployment, which impacts the economic viability of running large-scale AI services and influences pricing models for developers using their APIs.
The insatiable demand for AI compute infrastructure is driving significant cost challenges for leading AI labs, impacting their profitability and strategic decisions.
Meta Acquires AI Coding Startup Conductor to Enhance Its AI Capabilities
AI Agents & Tooling · TechCrunch · 2026-11-05
Meta has acquired Conductor, an AI coding startup founded in 2024 that launched its core product last summer. Conductor's platform allows users to spin up software projects by deploying multiple AI agents into parallel workspaces. These agents, from major labs like Anthropic and OpenAI, help software engineers work on compartmentalized parts of a project simultaneously.
Conductor's founder and CEO, Charlie Holtz, stated the company has "big, crazy plans for the next interface" and will have "something new" for customers by early next year. Meta confirmed the acquisition but provided no further details. Conductor's cloud service will shut down on November 20, with user cloud data deleted thereafter. This acquisition follows Meta's recent efforts to catch up in AI services, including the launch of its own Muse Code coding agent and the Moltbook viral agent platform acquisition.
Key facts
- Meta acquired Conductor, an AI coding startup founded in 2024.
- Conductor's platform uses AI agents from labs like Anthropic and OpenAI.
- Conductor's cloud service will shut down on November 20.
- Meta previously launched Muse Code and acquired Moltbook.
Why it matters: This acquisition could streamline Meta's internal software development workflows and potentially lead to new AI-powered developer tools that backend and AI engineers could leverage for faster project initiation and parallel development.
Meta is investing in AI-driven agentic development to accelerate its software engineering and AI product roadmap.
Groundcover Releases Kubernetes Rightsizing Beta from Wand Acquisition
Cloud & Infrastructure · Siliconangle · 2026-11-03
Observability company Groundcover Inc. has released a beta for its Autonomous Infrastructure feature, which automates the rightsizing of processor and memory requests for Kubernetes workloads. This technology is derived from Groundcover’s September acquisition of Kubernetes resource optimization startup Wand Cloud Ltd. The feature uses historical production data collected by Groundcover’s sensor to adjust resource requests, aiming to prevent out-of-memory errors from under-provisioning and reduce cloud costs from over-provisioning.
The beta allows customers on supported versions to immediately access the feature without new installations. It displays requested versus actual CPU and memory usage, estimated monthly savings, and tracks realized savings. Pricing calculations currently rely on customer-entered unit costs, with public cloud pricing and AWS Cost and Usage Report integration planned. Unlike tools that size workloads in isolation, Groundcover’s software prioritizes stability, checking node packing before increasing requests to avoid resource starvation. The feature is opt-in, allowing engineers to select deployments, namespaces, or clusters to manage, and it includes minimum/maximum bounds and an audit log. It is available at no extra cost to Groundcover customers running its sensor on Kubernetes.
Key facts
- Groundcover launched Autonomous Infrastructure in beta.
- Technology acquired from Wand Cloud Ltd. in September.
- Automatically adjusts Kubernetes processor and memory requests.
- Available to existing Groundcover sensor customers at no additional cost.
Why it matters: This tool can significantly reduce cloud spend and improve Kubernetes cluster stability by automatically optimizing resource allocation, freeing up backend and systems engineers from manual tuning and troubleshooting resource-related issues.
Automated Kubernetes rightsizing, now in beta from Groundcover, promises to optimize cloud resource utilization and reduce operational overhead.
Quick hits
- Discovered Materials Launches $1M Challenge for AI to Solve Materials Science Problems (TechCrunch) — Discovered Materials, a Y Combinator startup, is offering $1 million in prizes for AI models to solve five critical, unsolved problems in materials science.
- Anthropic Develops Method to Authenticate AI Model Responses with High Compute Cost (The New Stack) — Anthropic has developed a method to cryptographically prove which AI model generated a response, but it comes with a significant compute cost.
What to watch
- Monitor the practical impact of Anthropic's compute costs on its service pricing and developer offerings, especially given their IPO trajectory.
- Observe how quickly organizations can adapt their security posture to counter AI-driven exploit generation, particularly for Linux systems, and the adoption rate of new defensive AI tools.
- Track the evolution of LLMJacking and the development of new detection and prevention mechanisms from cloud providers and security vendors.
Sources
- OpenAI Pipeline Generates 948 Linux Root Escalation Exploits in 30 Hours for $1.2M — The Hacker News, 2026-11-05
- Google Cloud Reports Rise of LLMJacking: Threat Actors Monetize Compromised LLM Resources — Google, 2026-11-03
- CISA Adds Exploited Apple Flaw CVE-2024-27834 to KEV Catalog Amid Active Exploitation — The Hacker News, 2026-11-05
- Kremlin-Linked Hackers Exploited Patched iOS and Android Flaws in Watering Hole Attacks — Ars Technica, 2026-11-05
- APT29 Uses Hijacked Web Analytics Script for iOS/Android Exploits in Poisoned Source Campaign — The Hacker News, 2026-11-04
- Anthropic's Compute Costs Rise to 58% of Revenue, Impacting Margins Ahead of IPO — TechCrunch, 2026-11-03
- Meta Acquires AI Coding Startup Conductor to Enhance Its AI Capabilities — TechCrunch, 2026-11-05
- Groundcover Releases Kubernetes Rightsizing Beta from Wand Acquisition — Siliconangle, 2026-11-03
- Discovered Materials Launches $1M Challenge for AI to Solve Materials Science Problems — TechCrunch, 2026-11-05
- Anthropic Develops Method to Authenticate AI Model Responses with High Compute Cost — The New Stack, 2026-11-05