OpenAI Patches Critical AI-Discovered Flaws

Key Takeaways

• OpenAI is devoting major resources to fixing AI-discovered security vulnerabilities found while scanning open-source software since May
• Node.js fixed a record 50 vulnerabilities this year, with PHP affecting nearly 70% of websites with a known server-side language
• Meta's share offering was oversubscribed five times over, with NVIDIA and AMD participating to fund AI facility investments
• Accountants now earn $60-$125+ per hour training AI models on platforms like Scale AI, Mercor, and DataAnnotation
• Samsung Electro-Mechanics reported 30% Q3 sales growth and 165% operating profit jump driven by AI datacenter demand
• Cisco AI Defense joined the ServiceNow Control Tower Alliance as a founding member to help customers govern AI safely
• Urban Outfitters is expanding AI tools through Bamboo Rose TotalPLM to estimate product costs earlier in the design process
• IndiaMART targets 10% revenue growth in FY27, with new AI-led products expected to help from Q4
• PIP-ToM achieved 92.56%-95.00% accuracy on tests measuring AI understanding of human perspectives using GPT-4
• DelSA outperformed existing methods on ChronoFake, a benchmark of 30,706 articles testing fake news with scrambled timelines

PIP-ToM improves AI understanding of human perspectives

Researchers developed PIP-ToM to help AI models better understand human beliefs and intentions. The tool focuses only on evidence that a specific person could have seen, avoiding irrelevant information. PIP-ToM uses GPT-4 and achieved accuracy rates of 92.56%, 95.00%, and 95.00% on three major tests. It outperforms existing methods in understanding mental states from an agent's point of view.

DelSA tackles fake news with scrambled timelines

Researchers created ChronoFake, a benchmark of 30,706 articles, to test fake news that mixes up the order of real events. They developed DelSA, a self-updating agent that selects the right evidence and checks timeline consistency. DelSA uses an LLM to analyze candidate evidence and a verifier-trained reranker to balance relevance and timing. It outperforms existing methods on ChronoFake and remains competitive on PolitiFact and GossipCop.

Tech companies struggle with AI-discovered security bugs

Tech companies are spending more time and resources fixing bugs found by artificial intelligence. Node.js confirmed and fixed a record 50 vulnerabilities this year. The vulnerabilities affect PHP, a programming language used by close to 70% of websites with a known server-side language. Workers like Collina are handling the surge seven days a week with AI agents helping overnight.

AI bug discoveries overwhelm Silicon Valley

OpenAI and other tech companies are racing to fix a flood of security vulnerabilities found by AI. On October 21, OpenAI leader Fouad Matin worked through the night patching two critical vulnerabilities. OpenAI began scanning open-source software for flaws in May and has found more than many researchers expected. The company is devoting significant computing power and multiple teams to cybersecurity efforts.

Meta share offering oversubscribed five times over

Meta Platforms received orders for five times the number of shares it offered in a new share sale. The offering aims to fund investment in AI facilities. NVIDIA and AMD also confirmed they would participate in the offering. The news highlights continued strong investor interest in AI development.

Accountants now train AI models for high pay

Accountants can earn $60 to $125+ an hour training AI models on platforms like Mercor, DataAnnotation, and Scale AI. They rate AI responses, write ideal answers, and stress test models using accounting and tax knowledge. A March 2026 benchmark found over half of model failures come from non-numeric reasoning errors. Licensed finance professionals are in demand because credentials matter for these expert roles.

Data discovery key to AI governance, experts say

Security teams must use data discovery and classification to control AI inside their organizations. Experts say enforcing policies is harder than writing them, so strong controls are needed. Companies should focus on access management and auditing overly permissive accounts, including those held by AI agents. Security leaders must support AI adoption with proper risk management instead of simply saying no.

Samsung Electro-Mechanics sees big AI-driven growth

Samsung Electro-Mechanics reported Q3 sales growth of 30% year over year to KRW 3,763.3B. Operating profit jumped 165% to KRW 691.0B, driven by AI datacenter demand in its Component and Package divisions. The company plans to focus in Q4 on expanding high-end AI, automotive, and optics supply.

Urban Outfitters expands AI tools for product design

Urban Outfitters Inc. is expanding its use of Bamboo Rose TotalPLM by adding the Create solution with VibeIQ and Decision Intelligence technology. Design and merchandising teams at its banners will see estimated product costs at the concept stage before samples are made. The company says earlier cost visibility helps reduce risky redesigns and speeds up development decisions.

Cisco AI Defense joins ServiceNow security alliance

Cisco AI Defense joined the ServiceNow Control Tower Alliance as a founding member to help customers govern AI safely. The integration lets Security teams discover AI assets, scan for malicious code, run red team tests, and apply runtime protection within ServiceNow. Two Cisco AI Defense integrations are already available on the ServiceNow Store for inventory and security exposure management.

IRON-Meme benchmark studies political memes and identity roles

Researchers created IRON-Meme, a benchmark of 1,110 political memes to study how identities and roles appear in online meme narratives. The framework covers social identities like gender and politics, viewpoints, and narrative roles such as moral character and power. Tests of nine vision-language models showed that large reasoning models performed best, but accuracy dropped for less common identity labels. The dataset and code are publicly available for future research.

IndiaMART targets 10% revenue growth in FY27 with AI products

IndiaMART InterMESH aims for about 10% revenue growth in FY27, with new AI-led products expected to help from the fourth quarter. CEO Dinesh Agarwal said the company is working to improve subscriber renewals, especially among silver-tier customers. Gold and platinum customers remain engaged, and the Busy Infotech business is growing at 25 to 30%. The company also reported a 16% year-on-year rise in deferred revenue and noted that inquiry volumes have fallen by about 10% from their peak.

Sources

• Who Knows What - and What Does It Support? Perspective-Bounded Evidence Calibration for Multimodal Theory of Mind
• Beyond the Evidence Coherence Trap: Agent-Driven Evidence Selection for Temporally Disordered Fake News Detection
• Tech Companies Battle Infestation of AI-Discovered Bugs
• AI Is Finding a Flood of Bugs, and Silicon Valley Is Rushing to Fix Them
• AI investment fervor persists: Orders for Meta offering are five times the shares on offer; NVIDIA and AMD participate
• Accountants are finding a new kind of work...Training AI models
• Data discovery and classification anchors AI governance
• Samsung Electro-Mechanics: AI datacenter demand drove 30% YoY sales growth and a 165% rise in operating profit
• Urban Outfitters Inc. boosts product development visibility with AI
• Cisco AI Defense joins ServiceNow's Control Tower Alliance
• IRON-Meme: A Benchmark for Structured Understanding of Identities and Roles in Online Meme Narratives
• IndiaMART targets 10% revenue growth for FY27, expects AI-led products to drive growth