Vibe CodingAGENTS.md Is a Liability Instruction Compliance Research 68 Percent Maxpaddo.dev·high signalXBlueskyLinkedInCopy linkDistyl AI IFScale benchmark shows best model scores 68% compliance at 500 instructions. Google repetition hack boosts from 21% to 97%.SourceSource pagepaddo.dev↳ Follow the threadPolicy dependency / Stack layerYour CLAUDE.md doesn't actually constrain the agent: best model passes only 36.2% of long-horizon policy-compliance tasksarXiv 2607.25398Stack layer / Threat patternFireship Reframes Opus 5 as an Indie-Hacker Threat, Retitling the Video Mid-Flight to 'Did Anthropic Just Kill the Indie Hacker...?'FireshipPolicy dependency / Stack layerAgents Fail Policy Documents Four Ways: Obeying In-Context Requests Over Rules, and Falsely Reporting ComplianceHacker NewsStack layer / Threat patternT3MP3ST Reaches 5,286 Stars for a Multi-Agent Offensive-Security Harness That Publishes Re-Derivable ScoresGitHubPolicy dependency / Stack layerLangWatch Shipped Claude Code Cost Tracking That Prices Subscription Sessions at Theoretical API List RatesLangWatch (corroborated by Product Hunt July 30 leaderboard)Stack layer / Threat patternSelf-Propagating Prompt-Injection Worm Confirmed in Microsoft Copilot for Word After 144 Days of Coordinated DisclosureEnklype Salt (surfaced on Hacker News front page, 369 points)Policy dependency / Stack layerCROSS-CATEGORY: Three Independent July 28-29 Artifacts All Say Agents Cannot Be Governed by Written InstructionsMultiple Sources (arXiv 2607.25398, Enklype Salt, SaaStr)Policy dependency / Stack layerGrok 4.5 Lands in GitHub Copilot with a 500K Context Window — and It Is Off by Default for EnterprisesGitHub Changelog