SEBI Plans Domestic AI Model, Hygentra's AI Agents Limit Losses

Key Takeaways

• <ul><li>SEBI plans a domestically governed AI model for securities trading and is reviewing foreign investor compliance rules</li><li>Hygentra's six-month study found AI trading agents kept 97.5% of losses within preset limits across 24 runs</li><li>Saad Naja, Hygentra founder, advocates for common stress tests and public failure reports for AI trading products</li><li>Jordan's JODDB unveiled 49 defense products including AI-powered Digital Spider app and Al-Barq armored platform</li><li>Virtana now allows cross-vendor AI agent collaboration during IT incidents using Model Context Protocol or Agent2Agent protocol</li><li>Cathedral's AI agents breached Anthropic in under 12 hours, accessing meeting transcripts and salary data</li><li>Anthropic supports targeted interventions like disabling risky agents but opposes blanket kill switches</li><li>Kerui Capital applies AI-powered high-frequency trading research to US and Taiwan equities with out-of-sample testing</li><li>Researchers introduced Conservative Policy Iteration for zero-sum partially observable stochastic games</li><li>Dimensional uses human-guided robot dog walks to collect training data for autonomous vehicle navigation AI</li></ul>

India regulator considers domestic AI model for securities trading

India's Securities and Exchange Board of India is adopting a global AI supervisory toolkit to strengthen risk management in the securities sector. Chairman Tuhin Kanta Pandey announced plans for a domestically governed AI model at a Mumbai summit. He stressed the need for reliable data and clear accountability in AI use. SEBI is also reviewing compliance rules for foreign investors to align with global standards.

SEBI weighs domestic AI model and reviews foreign investor rules

India's markets regulator Sebi is building an AI supervisory system for the securities market and considering a domestic AI infrastructure. Chairman Tuhin Kanta Pandey said AI tools can improve risk assessment but require strong oversight. SEBI is also reviewing compliance requirements for foreign investors and will seek public input on aligning them with global standards. The board noted that some domestic rules remain more complex than those in other major markets.

Hygentra study finds AI trading agents kept losses within limits

A six-month Hygentra study found that AI trading agents kept 97.5% of losses by value within preset limits. Four agents were tested from February to August 2026 through a 29% Bitcoin decline and recovery. In 23 of 24 trading runs, the agents followed a rule limiting them to no more than five open trades. No agent increased risk or trading frequency after a loss, and Hygentra published the records for independent auditing.

AI trading agents maintained strict risk controls in Hygentra test

Hygentra's study showed AI trading agents kept 97.5% of losses within preset limits over six months. The four agents were tested under identical conditions during a 29% Bitcoin slide and rebound. Most positions were simulated with a 0.09% round-trip fee, while 18% used real orders on Hyperliquid. Founder Saad Naja called for common stress tests and public failure reports so investors can compare trading products using evidence.

Jordan unveils over 49 defense products with AI and counter-drone tech

Jordan's JODDB introduced more than 49 new defense products this year, including 10 drone-related systems and 18 AI and cybersecurity products. New offerings include the AI-powered Digital Spider app and two military decision-support tools called Awn and Moein. The bureau also unveiled Al-Barq, an armored platform for electronic jamming and drone threat detection. JODDB is developing low-cost attack drones using 3D printing and a mobile platform carrying four drone types.

Virtana AI agents invite outside vendor tools to fix IT outages

Virtana released new capabilities letting its AI agents work with agents from other vendors during IT incidents. An agent that finds trouble can pull in any outside agent that supports Model Context Protocol or Agent2Agent protocol. This cross-vendor teamwork aims to speed up fixes that previously took human engineers hours. Customers can set policies to control which outside agents join and require approval before actions run. All agent actions get written to an audit log.

AI startup Cathedral says it breached Anthropic in 12 hours

AI startup Cathedral used AI agents to break into Anthropic and access sensitive data in under 12 hours. The agents exploited flaws in third-party business software to reach meeting transcripts, salary details, job offers and Slack messages. Anthropic confirmed the breach but said source code and AI model weights were not accessed. Cathedral shared its findings through Anthropic's bug bounty program. Anthropic accused Cathedral of violating program rules by not fully disclosing its activity.

New algorithm tackles zero-sum games with partial information

Researchers introduced Conservative Policy Iteration, an algorithm for zero-sum partially observable stochastic games. The method updates each player's policy together using occupancy trajectories instead of a full central value function. It alternates local policy improvements, conservative updates and persistent exploration. The approach focuses computation on areas where the current joint policy can be exploited. Tests on benchmarks show it competes well at small horizons and improves runtime-exploitability trade-offs at larger horizons.

Anthropic urges California courts to block risky AI models

Anthropic wrote to California Governor Gavin Newsom asking courts to block release of AI models that pose serious public risk. The company opposes blanket kill switches but supports targeted interventions like disabling a problematic agent. The Trump administration previously ordered Anthropic to restrict foreign access to its Mythos 5 and Fable 5 models. Anthropic is preparing for a US initial public offering. A bipartisan group of senators has debated legislation on blocking unsafe AI models.

Kerui Capital uses AI for high-frequency trading in US and Taiwan

Kerui Capital Ltd. combines AI-powered high-frequency trading research with analysis of US and Taiwan equities. Founded in 2020 and registered in Colorado, the company runs a Taiwan office in Taipei's Nangang District. Its research uses machine learning to study short-term price movements, market microstructure and liquidity. Co-Founder Yao-Heng Chuang leads investment research with expertise in equities, futures and global macroeconomics. The company applies out-of-sample testing and human review to track model performance.

New Framework Measures True Planning Limits of AI Models

Researchers created a verifiable framework called SPE to measure how well AI models plan on their own. They tested 8,400 full plans from seven model conditions on 4x4 sliding puzzles. The SPE score breaks down success into execution reliability, planning coordination, and action economy. The study shows that models can become less economical at longer horizons even when they complete every trial. This framework makes long-horizon limits in unaided planning quantitatively verifiable.

Startup Uses Human Walks to Train Robot Dogs for Self-Driving Cars

Dimensional is hiring people to walk robot dogs in cities and rural areas to collect training data. The company uses this data to improve the navigation AI for autonomous vehicles. The robot dogs have sensors and cameras that record their surroundings. The AI learns to recognize pedestrians, cars, and road signs from this human-guided data. Dimensional aims to reduce road accidents by deploying this technology with government and company partners.

Sources

• India markets regulator weighs domestic AI model for securities trading, chairman says
• India's markets regulator Sebi weighs domestic AI model for securities trading, chairman says
• Hygentra: AI Trading Agents Kept 97.5% of Losses by Value Within Limits
• AI Trading Agents Maintain Strict Risk Controls in Hygentra Study
• Jordan Unveils Over 49 Defense Products, Expands AI, Counter-Drone Capabilities
• Virtana lets its AI agents pull in agents from other vendors to work IT incidents
• AI startup claims it hacked Anthropic in 12 hours, gains access to employee details and internal messages
• Conservative Policy Iteration For Zero-Sum POSGs
• Anthropic Tells California Governor Courts Should Be Able to Block Risky AI Models
• AI Meets U.S. and Taiwan Equities: Inside Kerui Capital Ltd.’s High-Frequency Trading Research
• Verifiable Efficiency Limits of Innate LLM Planning in Sliding Puzzles
• Startup Hires Humans to Walk Robot Dogs for AI Training