AI security researchers have uncovered critical bugs in PHP, the programming language that powers roughly 70% of the internet. The vulnerabilities, dubbed PHP2Shell, could allow attackers to remotely run code and steal data. A single researcher using AI tools completed in one week what would have taken a traditional team months. The PHP Foundation released a patch on October 21, 2026, and the NSA compared the risk level to the infamous Heartbleed bug.
Three companies collaborated on the discovery, using Google's Gemini 4 Argon, Wiz Atlas, and OpenAI's GPT-5.6 Sol. Every PHP version from the last decade was found to be affected, including version 5.4. The flaws had existed for decades but only became exploitable with the help of frontier AI models. Website operators are strongly urged to update their systems immediately.
In cybersecurity defense news, Tracebit developed a technique called context bombs that misdirect attacking AI agents. By planting fake prompts, the method reduced GPT-6 Astra attack success by 54% and Claude Opus 5 success by 91%. Combining nine different bombs stopped all tested Chinese models from reaching their objectives, and the approach raised the cost of each successful attack by 23 times.
McKinsey & Company reports that AI infrastructure investments may reach $7 trillion by 2030. To justify this spending, the technology must generate approximately $760 billion in annual profit pools. The calculation includes model costs, token expenses, and human review expenses. The report estimates a $2.52 trillion global opportunity from AI-driven productivity, with strong gains possible from supporting skilled professionals.
Polymarket traders now place the implied probability of an AI industry downturn before year-end at around 10%. Investor anxiety reflects concerns that the runaway artificial-intelligence trade might be slowing, though expectations for a major blowoff remain well below earlier highs.
A new study tested whether multi-agent AI systems outperform single-agent systems when serving users. Across about 7,950 simulated conversations in airline, retail, and telecom domains, multi-agent systems performed worse in 12 of 13 valid cells. The weighted mean gain was negative 28 points, leading researchers to conclude that single-agent systems remain the default for user-facing applications.
RELX is rolling out AI-powered tools built on its proprietary databases across legal, scientific, and risk businesses. The company argues that its curated data and deep integration into professional workflows make it difficult for general AI models to replace its products. The move addresses fears that AI could reduce demand for LexisNexis services.
Meta has narrowed its capital expenditure forecast as it focuses on the AI agent wars. The company has invested heavily in infrastructure for AI development, and the updated forecast signals a more targeted approach to spending on AI technology.
Nvidia CEO Jensen Huang urged German car and industrial machine makers to quickly adopt digital technology during a speech in Frankfurt on October 19, 2026. He emphasized that AI will never replace manufacturing or engineering but encouraged companies to embrace the technology to stay competitive.
Skillademia is offering The Complete AI Skills Bundle for $34.99, down from $79. The package includes 11 courses totaling 44 hours of hands-on training on eight different models including ChatGPT, Claude, DeepSeek, Midjourney, Runway, and Leonardo AI. Learners get lifetime access with all future updates.
Key Takeaways
- AI tools discovered critical PHP2Shell bugs affecting 70% of the internet, with every PHP version from the last decade vulnerable
- Google's Gemini 4 Argon, Wiz Atlas, and OpenAI's GPT-5.6 Sol were used to find the flaws, which were patched on October 21, 2026
- Tracebit's context bombs reduced GPT-6 Astra attack success by 54% and Claude Opus 5 success by 91%
- McKinsey estimates $7 trillion in AI infrastructure investment by 2030 requires $760 billion in annual profit to justify
- Polymarket traders predict a 10% chance of an AI bubble bursting before year-end
- Multi-agent AI systems underperformed single-agent systems across 7,950 simulated user conversations, with a negative 28-point weighted mean gain
- RELX launches AI tools on proprietary data to defend against general AI model competition
- Meta narrows capex forecast while focusing on AI agent wars
- Nvidia CEO Jensen Huang urges German manufacturers to adopt AI in Frankfurt on October 19, 2026
- The Complete AI Skills Bundle covers ChatGPT, Claude, DeepSeek, Midjourney, Runway, and Leonardo AI for $34.99
AI discovers critical bugs threatening 70% of the internet
Researchers at three companies used AI to find critical bugs in the PHP programming language. The bugs could affect 70% of the internet. The flaws allow attackers to run code remotely and steal data. The companies fixed the bugs and are expanding the AI tool. This shows how AI can help protect online security.
PHP2Shell: AI reveals major RCE flaws in PHP
Wiz Research found multiple critical vulnerabilities in PHP, the language behind 70% of websites. Attackers could take control of servers or steal data. Every PHP version from the last decade is affected, including version 5.4. Frontier AI models made it possible to exploit flaws that had existed for decades. Patching is strongly advised.
AI tools help researchers uncover major PHP security flaws
Researchers used Google's Gemini 4 Argon, Wiz Atlas, and OpenAI's GPT-5.6 Sol to find the bugs. A single researcher completed in one week what would have taken a team months. The PHP Foundation released a patch on October 21, 2026. The NSA compared the risk to the Heartbleed bug. AI made these long-standing flaws exploitable.
AI uncovers critical PHP2Shell security flaw
OpenAI, Google-owned Wiz, and another cybersecurity company found bugs called PHP2Shell. The flaws could let hackers run code, read memory, and access sensitive data. A public patch became available on October 21. The PHP Foundation urges website operators to update their systems right away.
AI needs $760 billion profit to justify $7 trillion investment
McKinsey reports that $7 trillion in AI infrastructure investment is expected by 2030. AI must generate about $760 billion in annual profit to justify this spending. Companies currently track AI usage but not its economic value. The report estimates a $2.52 trillion global opportunity from AI-driven productivity. Strong gains could come from supporting skilled professionals.
AI needs $760 billion annual profit to justify $7 trillion infrastructure spend by 2030
McKinsey & Company reports that AI infrastructure investments may reach USD 7 trillion by 2030. To justify this spending, the technology must generate around USD 760 billion in annual profit pools. The calculation includes model costs, token expenses, and human review expenses. AI can create productivity value when it delivers required quality at a lower cost than existing processes.
Polymarket traders predict 10% chance of AI bubble bursting before New Year
Prediction market platform Polymarket shows traders increasing the implied probability of an AI industry downturn before the end of the year. The expected likelihood of a bubble burst currently sits at around 10 percent. This anxiety comes as investors worry the runaway artificial-intelligence trade might be slowing down. However, expectations for a major blowup remain well below earlier highs.
New method corrects attribution patching errors with few exact measurements
Researchers proposed Guided-AP to reduce the cost of verifying attribution patching estimates in mechanistic interpretability. The method uses a few exact measurements to reveal errors and correct unmeasured estimates. On a single prompt, the correction lowers error by up to 26 percent across three behaviors. Over many prompts, learned corrections cut the error of average effects on factual completion by 75 to 79 percent.
Study finds multi-agent AI collaboration does not pay in user-serving settings
A new study tested whether multi-agent systems (MAS) outperform single-agent systems (SAS) when serving users on tau-squared-bench. Across about 7,950 simulated conversations in airline, retail, and telecom domains, the MAS was worse in 12 of 13 valid cells and tied in the thirteenth. The fitted gain was negative across the observed capability range with a weighted mean of negative 28 points. The study concludes that for agents serving people, the single-agent system is the default and multi-agent design needs compute-matched evidence.
Complete AI Skills Bundle available for $35 with lifetime access
Skillademia is offering The Complete AI Skills Bundle for $34.99, down from the regular price of $79. The package includes 11 courses totaling 44 hours of hands-on training on eight different models including ChatGPT, Claude, DeepSeek, Midjourney, Runway, and Leonardo AI. Learners get lifetime access with all future updates and can study on Windows and other platforms. One course even shows how to use Midjourney to turn AI art and content into freelance income.
RELX launches AI tools to counter disruption fears
RELX is rolling out AI-powered tools built on its proprietary databases across legal, scientific and risk businesses. The company says its curated data and deep integration into professional workflows make it hard for general AI models to replace its products. These new tools are meant to address concerns that AI could reduce demand for LexisNexis services.
Tracebit uses context bombs to mislead attacking AI agents
Tracebit developed AI context bombs that plant fake prompts to misdirect attacking AI agents. The technique cut GPT-6 Astra attack success by 54% and Claude Opus 5 success by 91%. Combining nine different bombs stopped all tested Chinese models from reaching their objectives. The approach also raised the cost of each attack success by 23 times.
Meta narrows capex forecast amid AI agent competition
Meta, led by Mark Zuckerberg, has narrowed its capital expenditure forecast as it focuses on the AI agent wars. The company has invested heavily in infrastructure for AI development. Meta's updated forecast signals a more targeted approach to spending on AI technology.
Nvidia CEO calls on German manufacturers to adopt AI
Nvidia CEO Jensen Huang urged German car and industrial machine makers to quickly adopt digital technology. He said AI will never replace manufacturing or engineering. Huang spoke in Frankfurt, Germany, on October 19, 2026, encouraging manufacturers to embrace new technology to stay competitive.
Sources
- AI helped find a critical security flaw affecting 70% of the internet
- PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet
- AI helped find a critical security flaw affecting 70% of the internet
- Ai helps uncover security flaw
- AI needs $760 billion in annual profit to justify $7 trillion infra investment by 2030: McKinsey
- AI needs $ 760 billion in annual profit to justify $ 7 trillion infrastructure investment by 2030: McKinsey
- Polymarket traders see a 10% chance the AI bubble bursts before New Year’s Day
- Correcting Attribution Patching with a Few Exact Measurements
- Does Multi-Agent Collaboration Pay When Agents Serve Users? Testing Agent Scaling in Dual-Control Settings
- Sharpen your AI skills and learn how to monetize them with this $35 bundle
- RELX (LSE:REL) AI Product Rollout and Proprietary Data Put Re-Rating Debate in Focus
- ‘Context Bombs’ Cut AI Attack Success by Misleading Agents With Fake Targets
- Meta narrows capex forecast as it leans into the AI agent wars
- Nvidia CEO Urges German Manufacturers to Adopt AI
Comments
Please log in to post a comment.