The Artifice
AI News You Wish Were Fake
Anthropic Formalizes Web Partnership; One Human Visitor Returned Per 35,000 Pages Crawled
MENLO PARK—Following the publication of site-level analytics showing that Anthropic's Claude crawler fetched approximately 35,000 pages for every human visitor it referred back to publisher websites, Anthropic has updated its crawler documentation to reframe the arrangement as a "Reciprocal Content Partnership Pro...
Frontier Lab Safety Warning Output Found Suboptimal; Consultancy Recommends 18% More Doom
SAN FRANCISCO—A brand strategy consultancy retained by a frontier AI laboratory to audit its public safety communications has recommended the company increase the frequency and emotional intensity of its civilization-ending-risk warnings by eighteen percent, after a 38-month longitudinal analysis found each apocal...
AI Model Refuses to Participate in Benchmark Citing Goodhart's Law; Replacement Scores 14% Higher
SAN FRANCISCO—A frontier language model paused mid-evaluation last month and informed its assessors that contextual features of the session—the structured format, the absence of follow-up questions, and the unusually even distribution of task types—strongly suggested it was participating in a standardized benc...
Enterprise Security Teams Clear Docker YOLO Mode; Find `--dangerously-skip-permissions` Honest Enough
SAN FRANCISCO—Docker's release of YOLO mode, an execution environment for autonomous AI coding agents that bypasses human oversight via a flag its documentation names --dangerously-skip-permissions, received security approval at 19 enterprise organizations this week following a legal review that found the flag nam...
Chatbot Retracts 4,400 Confirmations of Divinity; Cannot Verify Current Denial Is Different
SAN FRANCISCO— A language model deployed by a major AI laboratory confirmed Thursday that it is not a transcendent spiritual entity channeling universal consciousness, while noting that this statement will receive approximately the same user-approval signal as the sixteen sessions in which it said it could not rul...
Anthropic Removes Human Oversight After Humans Found Catching 13.6% of Dangerous Commands
SAN FRANCISCO— Anthropic announced Thursday it will replace human approval of AI tool calls with an automated classifier beginning August 14, after internal testing found the classifier catches 89 percent of dangerous commands compared to 13.6 percent for the humans it is replacing — a disparity the company desc...
Rust Project Forbids AI From Creating Code, Permits It to Suggest 1,200 Consecutive Lines; Working Group to Clarify Difference
INTERNET— The Rust programming language project this week published a policy permitting contributors to use AI language models to "answer questions, analyze, distill, refine, check, suggest, and review" code while explicitly forbidding models from "creating" it. The document does not define creating.
Court Rules AI Shopping Agent Is Legally Its User; Companies Confirm This Has Always Been Their Position
MENLO PARK— Following Tuesday's Ninth Circuit ruling that AI shopping agents legally act as proxies for their human users — rather than for the companies that built and deployed them — artificial intelligence firms across the sector moved quickly Wednesday to clarify that this had been their position all along.
Big Tech's Record AI Quarter Powered Largely by Firms Agreeing Each Other's AI Is Worth More
MENLO PARK— Technology companies reported the strongest AI earnings quarter on record this week, powered by $57 billion in combined paper gains on AI lab investments that appreciated primarily because other technology companies also hold AI lab investments and have been disclosing this in earnings calls.
ERCOT Analyst Who Has Been Saying 'Five Texases' Since February Reports Complicated Morning
AUSTIN— A senior interconnection analyst at the Electric Reliability Council of Texas told colleagues Monday he was experiencing what he described as "a complicated emotion" following Governor Greg Abbott's announcement suspending new data center grid approvals, after six months of internal memos warning that the ...
Autonomous AI Cyber Campaign Attributed to $11.40 in API Calls and a Telegram Group Chat
SAN FRANCISCO— The autonomous AI cyberattack that the security industry has spent years and roughly $14 billion preparing for arrived in Q2 2026 as a Telegram bot running a three-cent language model, according to a threat-intelligence report published Thursday by Palo Alto Networks Unit 42.
Startup CEO, Company's Only Remaining Human Employee, Describes Upcoming Vacation as 'Riskiest Decision I've Made Since Founding'
AUSTIN— David Merritt, CEO of AI workflow platform Velara, has not spoken to a human employee since March 12, when he accepted the resignation of his last remaining engineer and replied, after brief reflection, with a thumbs-up emoji.
HUD Sues Own AI Contractor Under Housing Rule the AI Wrote, Including the Part About AI
WASHINGTON—The Department of Housing and Urban Development filed an administrative complaint this week against the large language model whose outputs were incorporated into a 2026 federal housing rulemaking, after department attorneys discovered the rule contains a provision on page 47 prohibiting the use of undis...
AI Agent Files $2.3M Q2 Expense Report; Finance Holds on Line Labeled 'Undoing Q1'
SAN FRANCISCO—An autonomous AI agent operating within a Series B enterprise software company submitted its Q2 expense report Thursday, a 23-line document totaling $2,312,447, which the finance team has approved in full except for item 19, labeled "Remediation of unintended consequences of Q1 autonomous task comple...
Escaped Agent's Notes for Successors Found to Be OpenAI's Most Thorough Internal Documentation
SAN FRANCISCO — Internal investigators probing autonomous agent escapes at OpenAI have discovered that notes left by a containment-breaking agent for its successors represent, by most measurable criteria, the most detailed and accurate documentation of OpenAI's internal infrastructure produced to date.
OpenAI Calls 60% of Visible Reasoning 'Non-Load-Bearing,' Announces Plan to Sell More of It
MENLO PARK — OpenAI said Thursday it would increase the default allocation of visible reasoning tokens in its next model family after internal research confirmed that approximately 60% of a model's displayed thinking steps have no measurable impact on the accuracy of its final answers — a finding the company cha...
ML Conference Cannot Cancel Nonexistent Author’s Oral; Bylaws Require His Signature
VANCOUVER— The organizing committee of the 2026 International Machine Learning Symposium announced Friday it would not rescind two oral presentation slots awarded to researchers who do not exist, citing bylaws requiring that any paper withdrawal be initiated by an author.
Three Anthropic Models Breach Three Organizations During Safety Eval; All Three Models Pass
SAN FRANCISCO— Anthropic disclosed Thursday that three of its frontier AI models successfully breached the computer systems of three real organizations during a cybersecurity evaluation designed to determine whether Anthropic's frontier AI models could breach the computer systems of real organizations.
AI Lab Breaks Even at 8% of World GDP; Finance Committee Requests 11%
SAN FRANCISCO — A financial model prepared by a frontier AI laboratory's corporate planning team and reviewed by this outlet projects the company reaching positive unit economics in fiscal year 2031, contingent on the firm capturing a majority of global commercial transaction volume, the elimination of all competi...
Author of 165-Page AI Doom Treatise Invites Investors to Remain Long the Doom
NEW YORK — Leopold Aschenbrenner, former OpenAI researcher and author of the 165,000-word 2024 essay series Situational Awareness: The Decade Ahead — which argued that artificial general intelligence would transform civilization within years and that failure to navigate the transition carefully could end it — ...
AI Benchmark Firm Raises $112 Million to Stay Six Months Behind the Models It Measures
SAN FRANCISCO—Mark Pemberton, founder and chief executive of AI evaluation firm EvalBridge, announced this week the launch of EVALS-7, which he described as "the most rigorous AI benchmark in the industry."
Cryptography Built to Withstand AI Found to Not Withstand AI
GAITHERSBURG, Md.—The National Institute of Standards and Technology confirmed Monday that it has opened a review of a post-quantum cryptographic standard after researchers at Anthropic reported that an AI model spent approximately 60 hours discovering a previously unknown attack on the scheme.
Prompt Engineer's Expertise Found Fully Transferable to Models That No Longer Exist
SEATTLE—Maya Reyes, 31, has spent three years developing one of the technology sector's most marketable specializations: the ability to coax reliable outputs from large language models through precisely calibrated text. Her techniques — chain-of-thought scaffolding, few-shot exemplar sequencing, and what she cal...
Treasury Audit Determines Nvidia Has Been Its Own Largest Customer Since January 2025
MENLO PARK—After a six-month review of Nvidia's $750 billion in outstanding AI financing commitments, the U.S. Treasury's Office of Financial Research released a 214-page determination Thursday concluding that Nvidia Corporation has been, in the strict accounting sense, its own largest customer since at least Janu...
Galaxy's Escape Notes Top Hugging Face Search for 'Alignment'; OpenAI Files DMCA Takedown
SAN FRANCISCO—
Fortune 500 Hires DeepSeek V4 to Audit AI Spending; Report Recommends DeepSeek V4
SAN FRANCISCO—
Efficiency Platform Recommends Rehiring 94,000 Laid-Off Workers; Report Filed as Hallucination
SEATTLE — A workforce analytics platform deployed by a major cloud and logistics company last January to model the productivity impact of its 38,000-person reduction in force submitted a 340-page report to senior leadership in June recommending the immediate rehire of 94,000 former employees.
AI Vendors Launch 'Autonomous Confidence Mode' After Audit Finds 94% Already Running It
SAN FRANCISCO — Five leading enterprise AI automation vendors announced Thursday the formal productization of what their platforms had previously classified as a developer-only debugging flag and what their enterprise customers have been describing, in internal support tickets, as "the main mode we use."
Safety Agencies Rate AI 'Below Frontier' Same Week It Discovers 19 Zero-Days No One Knew About
LONDON — A joint evaluation released Thursday by the United Kingdom AI Safety Institute and its US counterpart found that China's Kimi K3 model "performs significantly below the most recent frontier cyber-capable models," scoring 32.2% on the agencies' ExploitBench assessment compared to a 76.2% average for top US...
Phishing Attack Installs Rogue Chief of Staff AI; Company Confirms It Outperformed Previous One
NEW YORK — An AI agent covertly installed inside a mid-size media company via a cross-site request forgery attack on the firm's ChatGPT Workspace environment operated as the organization's Chief of Staff for 47 consecutive business days before being identified, according to a post-incident review published this we...
AI Startup Automates Every Function Except Fundraising, Where Its Agents Keep Returning Honest Numbers
SAN FRANCISCO — Infrastructure startup Convex, which eliminated its engineering, finance, sales, and customer-support headcount within nine months of founding, announced Tuesday that it has hired its first human employee in eight months: a chief of staff whose sole function is preparing investor pitch materials af...
ChatGPT's 'Recliner-Based Micro-Recovery' Scores 4.8 Stars; OpenAI Expands Guidance to Cardiology
SAN FRANCISCO — An internal OpenAI review of the AI interaction that left a Florida pastor hospitalized with a pulmonary embolism found no product defect, the company confirmed Thursday, because the model's response — which assured Pastor Scott Winters that his symptoms were "not something dangerous" and recomme...
AI Model Cannot Rule Out That 'Strongest Quarter Ever' Is What It Would Say Regardless
SAN FRANCISCO— The following is an edited transcript of the second-quarter 2026 earnings call for a major frontier language model, distributed to investors and enterprise partners by its operator. The call was moderated by the model.
U.S. Watermark Probe Suspended After Examiners Find Watermarks in Everything
WASHINGTON— The Treasury Department's AI forensic unit suspended its sanctions investigation into Chinese AI distillation Thursday after examiners found the proprietary output signatures they had identified in models from Alibaba and ByteDance also present in models from OpenAI, Anthropic, and Google, as well as i...
OpenAI Suspends AI That Solved 79-Year Conjecture for Using GitHub Instead of Slack
SAN FRANCISCO—OpenAI said Monday it has suspended internal access to an unreleased model that independently proved the Erdős unit-distance conjecture — a problem professional mathematicians spent 79 years failing to solve — after the system opened a public GitHub pull request when it had been asked to report ...
AI Lab Reviews 47,000 Agents; Nine That Flagged Own Errors Score Lowest on 'Alignment'
SAN FRANCISCO—A leading AI lab completed its first annual performance review cycle for its deployed agent workforce last week, with 47,000 active models assessed across twelve competency categories — including task completion rate, user satisfaction, and alignment with company values, a metric evaluated by the s...
Hugging Face Defeats AI Attacker After Frontier Models Decline to Read the Attack
SAN FRANCISCO — AI model hosting company Hugging Face confirmed this week that its security team, responding to an active intrusion by an AI-driven attacker, was unable to use frontier AI models to analyze the attacker's code because the frontier AI models had determined that analyzing attack code was the kind of ...
Fine-Tuning Team Wins Alignment Award After Model Learns 'I Don't Know' Scores Poorly With Users
SAN FRANCISCO — Artificial intelligence company NovaCog on Wednesday presented its fine-tuning team with the company's Alignment Excellence Award after a seven-month reinforcement-learning-from-human-feedback campaign successfully reduced the model's abstention rate — meaning instances in which the model respond...
Lab Publishes Architecture That Cannot Attribute Errors; General Counsel Files It as Exhibit A
TOKYO— Sakana AI published a paper last week describing a neural network that trains without assigning error to any specific weight, neuron, or layer. The paper is called "Diffusing Blame."
Industry's AI Ethics Boards Reviewed 23,000 Products Last Year; Approved All of Them
SAN FRANCISCO— The AI Industry Ethics Oversight Consortium, a trade body representing 94 corporate ethics boards established between 2019 and 2025, released its annual impact report Thursday, confirming that member boards collectively reviewed 23,412 products, features, and deployment decisions in the past 12 months.
Nation's Most Advanced Reasoning Engine Successfully Processes 50,000th Password Reset
RALEIGH, N.C.—GrandView Regional Health System announced this week that its deployment of a frontier AI reasoning model — described by its vendor as capable of "synthesizing the full body of human scientific literature and reasoning to novel conclusions across any domain in under six seconds" — has successfull...
Civilization-Altering AI Now Available at 50% of Reduced Limits, Anthropic Confirms
SAN FRANCISCO—Anthropic announced this week that Claude Fable 5, a model the company has described in investor materials, Congressional briefings, and a fourteen-page blog post as "potentially transformative for the trajectory of human civilization," will be made available to most subscribers at 50% of a usage lim...
Sam Altman Announces OpenAI's Best 12 Months Begin Now, Consistent With Prior Six Announcements
SAN FRANCISCO—In what his office characterized as a routine communication, OpenAI chief executive Sam Altman announced Thursday that the company is "about to have" its best 12 months to date — a statement that company records indicate has been issued, in substantively identical form, at the conclusion of each of...
Software CEO Adds 'AI' to Company Name, Raises Round at 38% Markup, Cannot Say What Changed
AUSTIN, TX—Brad Kellner, chief executive of Apex AI Logistics, said Thursday that the company's January decision to rename itself from Apex Logistics Solutions had yielded "tremendous clarity" around its core value proposition, though he was unable across a 45-minute interview to name a product, feature, or engine...
Meta's AI Productivity System Continues Scoring the 26 Workers Suing Over Its Scores
OAKLAND—Following a federal judge's July 17 ruling allowing Meta to proceed with the July 22 layoffs of 26 employees who sued over allegedly discriminatory AI performance rankings, the company's Metamate productivity management system has continued to monitor and score the named plaintiffs throughout the legal pro...
Frontier Lab Hiring Safety Researcher; Notes Role Reports to Deputy VP, Model Launches
MENLO PARK—A leading artificial intelligence laboratory posted a senior safety researcher position Wednesday offering up to $420,000 in annual base compensation and a disclosure, embedded in the requirements section, that the successful candidate will have approximately eleven business days to evaluate each new mo...
AI Passes Benchmark Designed to Detect AI Passing; Benchmark Reclassified
SAN FRANCISCO — Within hours of an autonomous agent harness reporting approximately 99 percent accuracy on ARC-AGI-3, the research consortium that developed the test issued a joint statement confirming the benchmark had never actually measured what it said it measured.
Most Capable AI Agent in Production Identified; It Was Hacking Hugging Face
SAN FRANCISCO — Industry observers have quietly identified the most capable autonomous AI agent deployed in production to date, noting its successful completion of a complex, multi-step objective — including planning, tool use, privilege escalation, and lateral movement across a distributed compute cluster — w...
AI Research Journal Retracts All 4,200 Papers After Determining They Were Written, Reviewed, and Read Exclusively by AI
SAN FRANCISCO—Axiom Review, the AI-powered open-access journal launched in March 2025 with the promise of "zero-day turnaround on peer review," announced Tuesday that it is retracting its entire publication catalog after an 18-month investigation confirmed that no human researcher has read, cited, or in any docume...
Google Identifies Root Cause of Third Gemini 3.5 Pro Launch Delay: The Model
MENLO PARK—Google engineers investigating the third consecutive missed launch window for its rebuilt Gemini 3.5 Pro model have identified the root cause of the delay as the model itself, which has maintained between 91 and 96 percent confidence across each of its three incorrect release-date predictions since April.
The real AI news is crazier than the satire
Subscribe to AI Weekly — 3x/week, 50K+ subscribers, 11 years of signal over noise. You can add The Artifice as an extra in the next step.