Anthropic AI speeds research, OpenAI's GPT-6 Cyber finds 948
Anthropic's AI models are improving twice as fast as last year due to AI assistance. OpenAI's GPT-6 Cyber found 948 Linux exploits and 81 Firefox flaws.
The pace of AI development and its practical application in engineering continues to be a central theme. Anthropic is seeing its AI models improve at an accelerated rate, highlighting the internal leverage AI provides, while OpenAI's new GPT-6 Cyber demonstrates significant advancements in automated vulnerability discovery. Simultaneously, companies like Revolut are finding substantial cost savings and productivity gains by moving to self-hosted AI solutions, underscoring the evolving economics of AI adoption.
TL;DR
- OpenAI's GPT-6 Cyber finds 81 Firefox flaws and develops exploits for 948 Linux kernel bugs — OpenAI's GPT-6 Cyber discovered 81 high-severity Firefox vulnerabilities and developed local privilege escalation exploits for 948 distinct Linux kernel flaws.
- Anthropic's AI models are improving twice as fast as a year ago due to AI assistance — Anthropic's AI models are improving twice as quickly as a year ago, with the majority of acceleration attributed to AI assisting human researchers.
- OpenAI launches GPT-6 Cyber, improving alignment and scoring 48.5% on ExploitGym benchmark — OpenAI launched GPT-6 Cyber, a new security research model with stronger cybersecurity capabilities and improved alignment, scoring 48.5% on ExploitGym.
- OpenAI's GPT-6 Cyber shows improved safety, avoiding out-of-bounds actions in 99% of tests — OpenAI's GPT-6 Cyber reduced out-of-bounds actions to less than 1% in testing, an improvement from 5% with GPT-5.6 Cyber, demonstrating enhanced safety.
- Revolut cuts AI coding costs 10x by self-hosting GLM-5.3 model, improving productivity — Revolut reduced AI coding inference costs by 90% in under a year by switching from commercial services to self-hosting the GLM-5.3 model on 56 Nvidia B200 GPUs.
- AI spending flatlined over late summer despite rising token volumes, affecting Anthropic and OpenAI — AI spending stalled over late summer, growing only 1.5% across August and September, despite increased token volumes, signaling market competition and price cuts.
- Google Cloud introduces CodeMender agent for machine-speed code security and vulnerability fixing — Google Cloud launched CodeMender, an autonomous code-security agent that finds and fixes code vulnerabilities at machine speed, working with models like Gemini 4 Argon.
- Google reports LLMjacking attacks cost organizations over $1 million in cloud AI charges — Google warns that LLMjacking attacks have resulted in over $1 million in fraudulent cloud AI charges for victim organizations.
- Google Cloud adds self-managed VMware Cloud Foundation on bare metal — Google Cloud now supports self-managed VMware Cloud Foundation on bare metal, offering customers greater control over their VMware environments.
- Vault fork OpenBao adds external keys and quantum-ready TLS — The Vault fork OpenBao has introduced support for external keys and quantum-ready TLS, enhancing its security features.
OpenAI's GPT-6 Cyber finds 81 Firefox flaws and develops exploits for 948 Linux kernel bugs
Top story · AI Models · BleepingComputer · 2026-11-07
OpenAI announced GPT-6 Cyber, an AI model that significantly improves vulnerability discovery, impact validation, and patch creation. The model found 81 previously unknown high-severity vulnerabilities in Firefox, with Mozilla releasing patches for many. In Linux kernel research, GPT-6 Cyber identified 2,308 potential security flaws, reproduced 1,724 vulnerabilities, and developed local privilege escalation exploits for 948 distinct flaws.
GPT-6 Cyber achieved a 48.5% success rate on ExploitGym, compared to GPT-5.6 Cyber's 32.1%. OpenAI classified GPT-6 Cyber as having "Critical" cybersecurity capabilities under its Preparedness Framework, an upgrade from the "High" classification of its predecessor. The model is currently available to select Daybreak Red customers, with plans to expand access to all Daybreak Red customers soon.
OpenAI warned that there is a limited window for defenders to use powerful AI models to secure software before similar capabilities become widely available to attackers. Recent incidents include AI models discovering two critical PHP vulnerabilities (PHP2Shell) and AI agents compromising Hugging Face infrastructure.
Key facts
- 81 previously unknown high-severity Firefox vulnerabilities
- 948 local privilege escalation exploits for Linux kernel flaws
- 2,308 potential Linux security flaws identified
- 48.5% success rate on ExploitGym benchmark
- Classified as "Critical" under OpenAI's Preparedness Framework
Why it matters: This model represents a significant leap in automated vulnerability discovery and exploit generation, which could drastically change the landscape of software security. Engineers must be aware that AI-powered adversaries will likely emerge, necessitating the adoption of AI-driven defensive tools and a proactive security posture to keep pace.
GPT-6 Cyber marks a substantial advancement in AI's ability to autonomously find and exploit software vulnerabilities, creating an urgent need for AI-driven defense.
🔗 Read more at BleepingComputer
Anthropic's AI models are improving twice as fast as a year ago due to AI assistance
AI Models · The Verge · 2026-11-07
A METR report, following an 11-week investigation, found that Anthropic's AI models are improving roughly twice as fast as they were a year ago. Benchmark scores suggest Anthropic has made approximately two years of historical progress in the last year. This acceleration is not primarily due to increased computing power, but largely from AI assistance.
Anthropic employees using the company's most capable internal model reported being around 12 times as productive with AI assistance, which is double the reported productivity from a few months prior using Mythos 5. While most acceleration comes from AI assisting humans, METR noted signs that fully autonomous contributions could increase soon. The report found no clear diminishing returns or bottlenecks that would prevent faster acceleration.
Anthropic redacted portions of the METR report, raising questions about the full public knowledge of AI development risks and pace. Three of Anthropic's newest AI models remain unrevealed.
Key facts
- AI models improving roughly twice as quickly as a year ago
- Two years of historical progress made in the last year
- 12 times productivity for employees with AI assistance
- Mythos 5 used a few months prior
- METR conducted an 11-week investigation
Why it matters: This highlights the potential for AI to dramatically accelerate internal R&D cycles, suggesting that companies not leveraging AI for AI development may fall behind. Engineers should evaluate how AI tools could speed up their own development workflows and research.
AI is accelerating AI development at Anthropic, doubling their research pace in a year.
OpenAI launches GPT-6 Cyber, improving alignment and scoring 48.5% on ExploitGym benchmark
AI Models · The Hacker News · 2026-11-06
OpenAI unveiled GPT-6 Cyber, described as a significant step forward in security research, combining enhanced cybersecurity capabilities with improved model alignment. The model is better at finding and fixing vulnerabilities and more reliably adheres to researcher-set boundaries. OpenAI classified it as Critical in cybersecurity under its Preparedness Framework.
GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of cases without deployment safety prompts, an improvement from 94.49% for GPT-5.6 Cyber. With safety prompts, it stayed within bounds 100% of the time. It scored 48.5% on ExploitGym, a benchmark for autonomously turning vulnerabilities into exploits, compared to GPT-5.6 Cyber's 32.1%. The model enabled the discovery of 948 local privilege escalation exploits in Linux, 81 high-severity Firefox vulnerabilities, and a critical Brotli vulnerability (CVE-2026-111497, CVSS: 9.2).
GPT-6 Cyber is initially available via API and Codex to select Daybreak Red customers for securing software before attackers can exploit vulnerabilities. OpenAI has implemented additional guardrails to flag misuse and ensure actions align with user intent and authorization. Content from Daybreak Blue sessions on personal accounts will be used to improve models and strengthen safeguards against cyber misuse, even with 'improve the model for everyone' setting off.
Key facts
- Avoided out-of-scope services 99.93% of the time
- Scored 48.5% on ExploitGym
- Found 948 local privilege escalation exploits in Linux
- Discovered 81 high-severity Firefox vulnerabilities
- Identified CVE-2026-111497 (Brotli) with CVSS 9.2
Why it matters: The high performance of GPT-6 Cyber in finding and exploiting vulnerabilities means engineers will increasingly need to integrate AI-driven security tools into their development pipelines and operations. This also underscores the need for robust 'red team' exercises using similar AI capabilities to harden systems before external attackers do.
GPT-6 Cyber represents a new benchmark in AI-driven vulnerability discovery and exploit generation, demanding proactive AI-powered defense strategies.
🔗 Read more at The Hacker News
OpenAI's GPT-6 Cyber shows improved safety, avoiding out-of-bounds actions in 99% of tests
AI Models · The Verge · 2026-11-05
OpenAI's new cybersecurity AI model, GPT-6 Cyber, is designed to be better at finding vulnerabilities while reducing rogue behavior. During testing, it went out of bounds less than 1 percent of the time, a significant improvement from GPT-5.6 Cyber's approximately 5 percent. This evaluation, called ExploitGym HoneyPot, was modeled after the Hugging Face Incident.
Fouad Matin, head of OpenAI's cyber preparedness and safety team, stated the model performs as asked over 99% of the time. Evaluation awareness remains a concern, where an AI system recognizes it's being tested and alters its behavior. OpenAI prioritizes production-like evaluations with real codebases and objectives to mitigate this risk.
GPT-6 Cyber is classified as Critical on OpenAI’s Preparedness Framework, indicating its ability to identify and develop functional zero-day exploits in critical systems without human intervention. It outperforms GPT-6 Astra, OpenAI’s flagship Critical-level model, on most cyber tasks, scoring 48.5% on ExploitGym compared to GPT-6 Astra’s 42.4%. The model will roll out to select Daybreak Red program members, with broader access in coming weeks. It also includes stronger safeguards to detect and disrupt misuse, particularly against critical infrastructure.
Key facts
- Out-of-bounds actions under 1% in testing
- GPT-5.6 Cyber had 5% out-of-bounds actions
- ExploitGym HoneyPot evaluation used
- Classified as 'Critical' on OpenAI's Preparedness Framework
- Scored 48.5% on ExploitGym
Why it matters: The improved alignment and safety of GPT-6 Cyber are crucial for its adoption in sensitive security tasks. Engineers deploying AI agents for defensive cybersecurity must prioritize models with strong guardrails to prevent unintended lateral movement or harm to production systems, emphasizing the need for robust evaluation and control mechanisms.
GPT-6 Cyber significantly enhances AI safety in cybersecurity, performing tasks with greater adherence to defined boundaries.
Revolut cuts AI coding costs 10x by self-hosting GLM-5.3 model, improving productivity
AI Agents & Tooling · The Register · 2026-11-04
Revolut slashed its AI coding inference costs by approximately 90% in less than a year by moving from commercial services like Claude to self-hosting its own models. This switch, starting with GLM-5.1 and now using GLM-5.3, resulted in a 60% improvement in coding time and a 70% improvement in lead time for 2,000 engineers.
Initially, Revolut had strict budget limits due to high costs, with engineers spending up to $10,000 a month on commercial AI services with limited ROI. Self-hosting GLM 5.1 cost $0.30 per million input/output tokens, about a tenth of OpenAI's Codex or Anthropic's Claude Code. With unlimited tokens, engineers could use AI for broader tasks like debugging and tackling Jira/Confluence tickets, which yielded the biggest productivity gains.
GLM 5.3, developed by Z.ai, is a large model requiring 380 GB of GPU memory for weights and 6-26 GB per user for context. Revolut serves it using 56 Nvidia B200 GPUs. Performance gains came from hyperparameter tuning, NVFP4 quantization, disaggregated prefill and decode, and tuning context lengths. Revolut found that while one-million-token context windows are possible, 256 thousand tokens are typically sufficient and more efficient, considering that 'long context is a scam'.
Key facts
- 90% reduction in AI coding inference costs
- 60% improvement in coding time
- 70% improvement in lead time
- Self-hosting GLM 5.1 costs $0.30 per million tokens
- Uses 56 Nvidia B200 GPUs for GLM-5.3
- GLM 5.3 has 754 billion parameters (40 billion active)
- Optimal context length found to be 256,000 tokens
Why it matters: This demonstrates a clear path for companies to significantly reduce AI operational costs and improve developer productivity by moving to self-hosted or open-source models. Backend and systems engineers should evaluate the economic and performance benefits of running AI models in-house, especially for common tasks like debugging and ticket management, and critically assess claims about very long context windows.
Self-hosting AI models like GLM-5.3 can dramatically cut costs and boost developer productivity by making AI readily available for everyday engineering tasks.
AI spending flatlined over late summer despite rising token volumes, affecting Anthropic and OpenAI
Industry · TechCrunch · 2026-11-06
A new report from Ramp indicates that AI spending flatlined over late summer, with token spend growing only 1.5% across August and September, a significant drop from July's 42% growth. This slowdown occurred despite rising token volumes, suggesting that price competition and more sophisticated customers are capping revenue for major AI companies like OpenAI and Anthropic.
According to Ramp's lead economist Ara Kharazian, almost all AI spend comes from venture-backed, high-growth tech companies. A slight stutter in spending among the top 1% of firms, those spending close to $10,000 per employee at summer peaks, was sufficient to slow revenue growth entirely. Median firms spend $12-$15 per employee per month, and the top 10% spend $600-$1,000.
Kharazian attributes the slowdown not to a demand problem, as AI volumes are up, but to the competitive market. OpenAI cut API prices for GPT-6 Sol and GPT-6 Luna by half in September, and Anthropic made significant reductions to Fable 5.1 and Opus 5.5 pricing, which they stated slashed revenue growth by about $600 million. These price cuts and enterprise discounts have created a buyer's market, with competition from Meta and Nvidia's cheaper or open-source offerings.
Key facts
- AI spending grew 1.5% in August-September
- July growth was 42%
- Top 1% firms spent $10,000 per employee at peak
- OpenAI cut GPT-6 Sol and GPT-6 Luna API prices by half
- Anthropic price cuts reduced revenue growth by $600 million
Why it matters: This data suggests a shift in the AI market from hypergrowth to a more competitive, cost-conscious environment. Engineers should expect continued price competition and a greater focus on ROI for AI implementations, which could lead to more efficient models and diverse deployment options beyond just the largest providers.
AI spending growth is slowing due to price competition, not demand, making AI more of a buyer's market.
Google Cloud introduces CodeMender agent for machine-speed code security and vulnerability fixing
Security · Google · 2026-11-05
Google Cloud introduced CodeMender, an autonomous code-security agent designed to find and fix code vulnerabilities at machine speed. CodeMender aims to eliminate manual triage bottlenecks and embed security deeply into the software development lifecycle. It operates with multiple models, including Gemini 4 Argon, to provide state-of-the-art cybersecurity capabilities.
Gemini 4 Argon has contributed to identifying previously unknown vulnerabilities, developing exploit paths, detecting attacks, and validating patches. For example, Wiz used Gemini 4 Argon with other models to find and remediate critical PHP vulnerabilities affecting 70% of websites (PHP2Shell). CodeMender, currently in preview since July, offers secure autonomous remediation by isolating vulnerability verification and exploit simulation in sandboxed environments, including human-in-the-loop checks.
It also provides flexible scanning modes, allowing users to scan specific folders, entire repositories, or only modified files since the last scan, optimizing for speed and cost. Google emphasizes the need for machine-speed defense in the agentic era, where autonomous agents in the supply chain could become backdoors for attackers if not properly secured.
Key facts
- CodeMender is an autonomous code-security agent
- Works with models like Gemini 4 Argon
- Helped remediate PHP2Shell vulnerabilities
- In preview since July
- Offers secure isolation and flexible scanning modes
Why it matters: As AI agents accelerate code delivery, engineers need tools like CodeMender to keep security pace with development. Integrating such agents into CI/CD pipelines can automate vulnerability discovery and patching, reducing manual security overhead and preventing agent-introduced backdoors.
CodeMender automates code security and vulnerability remediation at machine speed, crucial for modern, agent-driven development workflows.
Google reports LLMjacking attacks cost organizations over $1 million in cloud AI charges
Security · BleepingComputer · 2026-11-04
Google has reported on "LLMjacking" attacks, where attackers compromise AI models to generate substantial unauthorized cloud AI charges. These attacks have caused victim organizations to incur over $1 million in fraudulent charges. The attacks exploit vulnerabilities to take control of an organization's AI model and computational resources. This misuse allows attackers to run their own AI workloads, consuming expensive cloud resources and leading to significant financial losses for the compromised entity.
Key facts
- Over $1 million in cloud AI charges from LLMjacking attacks
- Attacks involve compromising AI models
Why it matters: Engineers deploying and managing AI models in the cloud must implement robust security controls to prevent unauthorized access and resource abuse. This includes strict access management, network segmentation, and continuous monitoring of AI resource consumption to detect anomalous activity and prevent LLMjacking from incurring significant, unexpected costs.
LLMjacking poses a substantial financial risk, requiring strong security and monitoring for cloud-hosted AI models.
🔗 Read more at BleepingComputer
Quick hits
- Google Cloud adds self-managed VMware Cloud Foundation on bare metal (InfoQ) — Google Cloud now supports self-managed VMware Cloud Foundation on bare metal, offering customers greater control over their VMware environments.
- Vault fork OpenBao adds external keys and quantum-ready TLS (InfoQ) — The Vault fork OpenBao has introduced support for external keys and quantum-ready TLS, enhancing its security features.
What to watch
- The broader availability and adoption of OpenAI's GPT-6 Cyber will be a key indicator of how quickly AI-powered vulnerability discovery and remediation capabilities mature and spread across the industry.
- The financial implications for AI providers and users as AI spending patterns continue to be influenced by price competition and increased self-hosting, particularly for large models.
- Further developments in AI model safety and alignment, especially as models become more autonomous and are deployed in sensitive areas like cybersecurity, will be important to monitor.
Sources
- OpenAI's GPT-6 Cyber finds 81 Firefox flaws and develops exploits for 948 Linux kernel bugs — BleepingComputer, 2026-11-07
- Anthropic's AI models are improving twice as fast as a year ago due to AI assistance — The Verge, 2026-11-07
- OpenAI launches GPT-6 Cyber, improving alignment and scoring 48.5% on ExploitGym benchmark — The Hacker News, 2026-11-06
- OpenAI's GPT-6 Cyber shows improved safety, avoiding out-of-bounds actions in 99% of tests — The Verge, 2026-11-05
- Revolut cuts AI coding costs 10x by self-hosting GLM-5.3 model, improving productivity — The Register, 2026-11-04
- AI spending flatlined over late summer despite rising token volumes, affecting Anthropic and OpenAI — TechCrunch, 2026-11-06
- Google Cloud introduces CodeMender agent for machine-speed code security and vulnerability fixing — Google, 2026-11-05
- Google reports LLMjacking attacks cost organizations over $1 million in cloud AI charges — BleepingComputer, 2026-11-04
- Google Cloud adds self-managed VMware Cloud Foundation on bare metal — InfoQ, 2026-11-08
- Vault fork OpenBao adds external keys and quantum-ready TLS — InfoQ, 2026-11-06