Anthropic is significantly expanding how vetted security professionals interact with its most powerful AI models. The company has merged its Project Glasswing and Cyber Verification Program into a single three-tier program. Defense Access offers the broadest reach, Red Team Access permits authorized penetration testing, and Specialized Access provides the fewest guardrails for testing critical infrastructure like power grids. Early results from CyScenarioBench showed the model completed 34 of 50 tasks in the Red Team tier without being blocked.
Key Takeaways
- <ul><li>Anthropic consolidated its cybersecurity AI programs into three tiers: Defense, Red Team, and Specialized Access</li><li>Project Glasswing partners found at least 129,000 verified software vulnerabilities between April and July 2026</li><li>Over 33,000 of those vulnerabilities were rated critical or high severity</li><li>SailPoint reports AI agent security spending reached $435 million between April and September 2026</li><li>86.9% of companies have delayed AI deployments due to governance gaps</li><li>ZoomInfo earned ISO/IEC 42001:2023 AI governance certification with zero findings</li><li>AWS published guidance for ISO/IEC 42005:2025 AI system impact assessments</li><li>Zetaris claims its cloud AI Data Harness can cut total cost of ownership by up to 67%</li><li>Microchip Technology shows 19% revenue growth and 41% EBIT growth in forward estimates</li><li>Microchip's PEG ratio of 0.49 represents a 62% discount to its semiconductor sector</li></ul>
Anthropic expands AI model access for security teams
Anthropic is widening its program that lets vetted cybersecurity experts test its strongest AI models with fewer safety limits. This follows Project Glasswing, which found over 100,000 software vulnerabilities this year. The company says it is committed to improving AI security and transparency.
Anthropic launches three-tier cyber AI access program
Anthropic merged its Project Glasswing and Cyber Verification Program into one expanded program with three tiers. Defense Access is the broadest tier, Red Team Access allows authorized penetration testing, and Specialized Access has the fewest limits for testing critical systems like power grids. Testing with CyScenarioBench showed the model completed 34 of 50 tasks in the Red Team tier without blocks.
Anthropic widens access to top AI models for cyber experts
Anthropic is expanding a program that lets vetted security professionals test its most powerful AI models with reduced safeguards. Partners in Project Glasswing found at least 129,000 verified software vulnerabilities between April and July 2026. Over 33,000 of those were rated critical or high severity. Anthropic says the real number is likely at least five times higher.
ZoomInfo earns ISO AI governance certification
ZoomInfo achieved ISO/IEC 42001:2023 certification for its Artificial Intelligence Management System after an independent audit by Schellman. The assessment found zero findings, zero nonconformities, and zero opportunities for improvement. The certification shows ZoomInfo governs AI effectively and will maintain it through periodic surveillance audits.
AWS guides customers on ISO AI governance standard
AWS published guidance to help organizations align with ISO/IEC 42005:2025 for AI system impact assessments. The standard provides templates and processes for evaluating AI risks, including how to scope assessments, integrate them into the AI lifecycle, and document outcomes. It also helps avoid duplication with other enterprise risk assessments.
SailPoint warns AI agent security lags adoption by 40 times
SailPoint says companies are adopting AI agents far faster than they are securing them. Between April and September 2026, $435M went into agent security funding. 86.9% of companies have delayed AI deployments due to governance gaps. SailPoint urges dynamic identity security instead of manual oversight.
Former Secret Service leader joins Bloo advisory board
Bloo appointed Michael H. Lashlee to its Advisory Board. He was previously Chief Security Officer at Mastercard and spent 25 years with the U.S. Secret Service. He will guide Bloo on enterprise security and telemetry for AI and operations. Bloo focuses on turning large data volumes into useful intelligence.
AI slop posts are hurting LinkedIn, content expert warns
Jasper Pye, who teaches AI writing, says generic AI posts are ruining LinkedIn. LinkedIn added a 'seems like AI slop' button on July 30. He says AI content often lacks a unique voice or real information. His advice is to use AI but keep a personal tone with original details.
Zetaris launches cloud AI data harness for secure access
Zetaris launched a cloud-based AI Data Harness that lets AI agents safely reach enterprise data in real time. It works across structured, unstructured and streaming sources without moving data. The company says it can cut total cost of ownership by up to 67 percent. A hackathon with SpaceXAI's Grok Build will test the system.
WarrenAI names top AI networking stocks for investors
WarrenAI listed Credo, Astera Labs and Marvell as top AI networking stocks. Credo leads with strong growth and a 67.1% gross margin. Astera Labs shows high growth but faces valuation concerns. Marvell offers broad custom silicon exposure with strong momentum but also downside risk based on fair value estimates.
Mexty launches full AI literacy program for UAE schools
Mexty introduced a complete AI Literacy curriculum for schools in the UAE. The program follows the UAE National AI Literacy Curriculum Framework for grades K through 12. It includes editable lessons, assessments, interactive activities, and AI Literacy Labs. Students learn to understand, question, verify, and use AI responsibly. Schools can edit all content to fit their own needs.
New method audits unreliable AI agent decisions
This research presents a way to check the output of black-box LLM agents. A low-cost open-weight surrogate model reads the same context and scores proposed actions using log-probabilities. The method can detect wrong argument values and holistically flawed calls without accessing the agent internals. On difficult coding tasks it reached an AUROC of 0.825, outperforming the actor's stated confidence. It also supports real-time review gates and confidence feedback.
Analysts highlight Microchip Technology as AI semiconductor pick
Steven Cress and Nicole Benjamin discussed Microchip Technology on Seeking Alpha's The Weekly Grade. The company is noted for strong growth compared with other semiconductor stocks. Forward estimates show 19 percent revenue growth, 41 percent EBIT growth, and 51 percent EPS growth. Its PEG ratio of 0.49 represents a 62 percent discount to the sector. The stock also scores well on momentum, profitability, and analyst revision grades.
Sources
- Anthropic opens its most powerful AI models to more security teams
- Anthropic is expanding verified AI model access to cybersecurity firms
- Anthropic opens its most powerful AI models to more security teams
- An outside audit found zero findings in ZoomInfo’s AI oversight
- Responsible AI governance: How AWS positions customers to align with ISO/IEC 42005:2025
- SailPoint’s 40x Gap — Enterprise AI Agent Adoption Has Outrun Identity Security by a Factor of 40
- Global Cybersecurity Leader Michael H. Lashlee Joins Bloo to Strengthen Enterprise AI Security
- I’ve spent a decade teaching AI to write. Here’s why your slop posts are ruining your LinkedIn
- Zetaris Launches Cloud-Based AI Data Harness, Giving AI Agents a Safe, Real-Time Path to Enterprise Data
- WarrenAI’s top AI networking stocks: Credo, Astera Labs and Marvell lead rankings
- Mexty Launches Complete AI Literacy Curriculum for UAE Schools
- Proxy Confidence: Auditing Black-Box LLM Agents with a Surrogate's Log-Probabilities
- Beyond Compute: The Unsung AI Semiconductor Worth Buying
Comments
Please log in to post a comment.