AdvertiseFee Deal Desk

Alphabet (Google)

Google Uncovers exploit_agent, First AI Hacking Tool With No Human in the Loop

Google's threat researchers found exploit_agent, an AI tool that exploited software flaws with no human help. Jason Wei has joined Mercor as chief scientist.

Be a creator
November 18, 2026, 07:52 AM UTC5 min read
AI SummaryAI
  • GTIG identified exploit_agent as the first AI tool to exploit software flaws with no human help.
  • MIDNIGHT NEPTUNE used AI in intrusions at crypto companies in North America and Europe.
  • Mercor closed a $750 million Series D at a $20 billion valuation, with NVIDIA investing.
  • Alphabet (GOOGL) traded at $282.69, down 0.62%, with strongest support at $272.58.

Google Logs First Fully Autonomous AI Exploit

Google's Threat Intelligence Group (GTIG) has identified the first AI tool that found software flaws and exploited them entirely on its own, with no human in the loop. The tool, named exploit_agent, sat on an attacker staging server left open to the internet, the group said in its third-quarter threat report published Monday by Alphabet's security unit. GTIG states that this is the first time it has seen attackers use large language model frameworks to automate flaw discovery and exploitation end to end.

The tool connected to several large language models, the AI systems behind chatbots, including leading commercial models and ones Google classifies as local. It watched widely used web software for patches, then compared the code before and after each fix to isolate the changed lines, which show where the flaw sits. Once the weakness was clear, exploit_agent drafted an exploit and ran it automatically. Comments inside the files, in Google's translation, read “LLM API keys for auto-exploitation (exploit_agent pilot)”.

The report shows the tool probing the access controls of a WordPress plugin called Spectra (Ultimate Addons for Gutenberg), version 2.19.25, but does not say whether that probing led to an attack. The same server held scripts for mapping a network after a break-in, guessing passwords through repeated attempts, and planting webshells for return access. Data taken from victims points to French websites across the media, entertainment and education sectors. Google also reports that overall disclosures and exploits nearly doubled between January and August 2026: zero-day attacks averaged eight a month in 2025, eleven a month in 2026, then jumped to 22 in August. The company adds that open-source attack frameworks CAIRN and PentAGI have appeared in real intrusions.

The report's North Korea section carries a crypto link: a cluster Google tracks as MIDNIGHT NEPTUNE used AI in intrusions at cryptocurrency companies in North America and Europe, posing as recruiters whose profile photos carried SynthID watermarks, Google's markers for AI-generated images. Once inside, the group altered a settings file in internal code repositories so that a backdoor called SOMBERMEME downloaded when developers opened the Claude Code coding assistant.

Jason Wei Leaves Meta for Mercor

AI researcher Jason Wei has joined Mercor, an AI recruiting and training-data platform, as chief scientist, a move he announced in a Wednesday post on X. Wei wrote that he had spent an exciting year in Meta's TBD Lab and that data will become a key frontier of AI development as more companies train their own models.

Wei is one of the most cited figures in language-model reasoning research. He joined Google Brain in 2020 as an AI resident, later working as a research engineer and research scientist. In 2022 he co-authored the paper that introduced chain-of-thought prompting, a method in which a model lays out intermediate reasoning steps before answering, lifting performance on math, common-sense and symbolic tasks. He also contributed to research on emergent abilities and instruction tuning. In 2023 he moved to OpenAI, where he worked on reasoning models and agents behind products such as the o1 model and Deep Research. In 2025 he left for Meta Superintelligence Labs and has now moved on after roughly a year.

Mercor was founded in 2023 as an AI-driven recruiting platform and has expanded into model training, expert data supply and model evaluation. It matches doctors, lawyers, engineers and finance professionals to projects that build training data, design evaluation questions and judge output quality, work general annotators cannot cover as AI moves into legal analysis, medical judgment and financial modeling. The company also builds reinforcement learning environments and agent evaluations that test multi-step tasks rather than text answers. Chief executive Brendan Foody said Wei, already a client and collaborator, will build the research team and lead data and evaluation work; Wei's benchmarks include SimpleQA, BrowseComp and HealthBench. The hire lands during fast expansion: Mercor just closed a $750 million Series D at a $20 billion valuation, with Benchmark, Felicis, Fundomo, NVIDIA and an Apollo-affiliated fund among the investors, about a year after a $350 million Series C at roughly $10 billion in October 2025.

GOOGL at $282.69 With Support at $272.58

For Alphabet, the two stories raise one market question: whether Google's security research and its pull on elite AI talent hold up while the stock drifts lower. COINOTAG data shows Alphabet (GOOGL) trading at $282.69, down 0.62% in the last 24 hours, inside a daily downtrend with an RSI of 37.8. COINOTAG's composite scoring puts the strongest support at $272.58, rated 86/100, and the strongest resistance at $291.62, rated 88/100; a daily close through either band would set the next leg. Positioning stays light: perpetual open interest of $177.4 million and a funding rate of 0.0128% point to no crowded leveraged book around the stock.

COINOTAG's editorial and research desk.

AI-Assisted

AI-generated, AI-reviewed, under COINOTAG editorial oversight.