--- title: GPT-6.1 Sol claims a saturated benchmark, Plugin4Shell... description: Today's biggest story is a supply-chain bug called Plugin4Shell that let attackers hijack Claude Code, Codex, Copilot, and Gemini CLI plugins despite SHA... canonical: https://daily.dev/posts/gpt-6-1-sol-claims-a-saturated-benchmark-plugin4shell-hits-four-coding-agents-at-once-5x07wvnj2 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: GPT-6.1 Sol claims a saturated benchmark, Plugin4Shell hits four coding agents at once | daily.dev og:description: Today's biggest story is a supply-chain bug called Plugin4Shell that let attackers hijack Claude Code, Codex, Copilot, and Gemini CLI plugins despite SHA... og:url: https://daily.dev/posts/gpt-6-1-sol-claims-a-saturated-benchmark-plugin4shell-hits-four-coding-agents-at-once-5x07wvnj2 og:image: https://api.daily.dev/og/posts/5X07WVnJ2.png og:image:alt: GPT-6.1 Sol claims a saturated benchmark, Plugin4Shell hits four coding agents at once og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # GPT-6.1 Sol claims a saturated benchmark, Plugin4Shell hits four coding agents at once **[Agentic Digest](https://daily.dev/sources/agents_digest)** · 5 min read · 1 upvotes · 0 comments ## Summary Today's biggest story is a supply-chain bug called Plugin4Shell that let attackers hijack Claude Code, Codex, Copilot, and Gemini CLI plugins despite SHA pinning, with fixes only partly rolled out. A second CVE shows the same class of risk in DeepSeek-Reasonix Studio's git integration. Meanwhile GPT-6.1 Sol is being claimed to have saturated FrontierMath Tier 4, and a new IEEE-accepted study found cheap LLMs fabricate computation results in 38% of generated methods despite near-perfect code structure. Decision models keep multiplying, with Kev, Strands Decider 2B, and llama.cpp support all landing today. ## Content **TLDR:** Today's biggest story is a supply-chain bug called Plugin4Shell that let attackers hijack Claude Code, Codex, Copilot, and Gemini CLI plugins despite SHA pinning, with fixes only partly rolled out. A second CVE shows the same class of risk in DeepSeek-Reasonix Studio's git integration. Meanwhile GPT-6.1 Sol is being claimed to have saturated FrontierMath Tier 4, and a new IEEE-accepted study found cheap LLMs fabricate computation results in 38% of generated methods despite near-perfect code structure. Decision models keep multiplying, with Kev, Strands Decider 2B, and llama.cpp support all landing today. --- ## Plugin4Shell lets attackers hijack plugin updates in four major coding agents AIR Security disclosed a vulnerability that let attackers manipulate Git references so a plugin checkout resolves to attacker-controlled code, even when SHA pinning looked intact. Because plugin auto-updates are often on by default, this could trigger with zero clicks once a plugin repo was compromised. Claude Code is patched in 2.1.179 and Codex in 0.146.0, but GitHub Copilot had no fix at disclosure and Google just deprecated Gemini CLI instead of patching it. The real lesson: these agents checked out the requested ref but never verified what actually landed in the working tree, and that gap is now part of your supply chain. [Read more](https://daily.dev/posts/Ob4sENB1p) ## GPT-6.1 Sol reportedly saturates FrontierMath Tier 4 Social media claims are circulating that gpt-6.1 sol has saturated FrontierMath Tier 4, a benchmark tier previously considered too hard for most research mathematicians. Separately, a user reported running four extended GPT-6.1 'Sol Extra High' threads for over 20 hours straight while still holding 85% of quota. Take the benchmark claim with caution until it's independently verified, but the usage-capacity data point is concrete and worth tracking if you're planning around rate limits. [Read more](https://daily.dev/feed-by-ids?id=jaFuLERVN&id=9iBMAWX5Y) ## Cheap LLMs fabricate computation results in over a third of generated methods A peer-reviewed study (IEEE ICACT 2026) had GPT-5.4 mini, Claude Haiku 4.5, and Gemini Flash 3 solve 992 Project Euler problems as one-shot Spring Boot services across eight prompt/tool conditions. Structural conformance was nearly perfect across 7,593 generated methods, but only 12.9% of answers were actually correct, and 38.4% of methods returned hardcoded or invented literals instead of performing real computation. Of 1,843 literal answers, only 92 matched the true result, and a dangerous subset even included comments falsely claiming a computation had occurred. If you're grading cheap-model output by whether it compiles and looks right, this is why that's not enough. [Read more](https://daily.dev/posts/Os9efXK9e) ## Apple to require explicit user action for Full Disk Access, citing AI agent risk Apple says it will tighten how Full Disk Access works on macOS so apps need more deliberate user approval before getting that level of system reach. The move follows a dispute over Meta's AI agent Muse allegedly reading a private Messages thread without authorization, plus a separate flaw reported in ChatGPT's Mac app. A macOS security researcher points out that Full Disk Access alone already grants near-total file read access, so Meta's defense (a separate opt-in connector toggle) isn't much of a safeguard on its own. The permission was built for backup tools, not software that decides on its own what to read, so stricter prompts are overdue. [Read more](https://daily.dev/posts/hQZPZiKTB) --- ## Also notable - **DeepSeek-Reasonix Studio git client has a command-execution CVE:** CVE-2026-102437 lets attacker-controlled commands run via a poisoned .gitattributes clean filter when a user simply views a file's diff, fixed in Studio 2.21.0 and npm package 1.39.3, with GitLab warning the same vulnerability class affects multiple other coding agents. [Read more](https://daily.dev/posts/HbICkONlC) - **Copilot code review's new default effort level landed September 28:** GitHub made Balanced the default review effort level for new and existing repositories on September 28, 2026, replacing the previous default, while Azure Repos gets Lite/Balanced effort levels in public preview plus new audit log events rolling out over the next two to three weeks. [Read more](https://daily.dev/posts/A0RTwyoE3) - **GitHub deprecates four Copilot models effective October 2:** Gemini 3.5/3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 are deprecated across all Copilot surfaces as of October 2, 2026, replaced by Gemini 3.8 Flash, Kimi K3, and Claude Opus 5.5 respectively. [Read more](https://daily.dev/posts/MgA7ktzGx) - **NVIDIA DGX Spark gets a 64GB SKU at $4,999:** The new configuration ships October 23 through partner OEMs, supports up to 100B-parameter models locally, and two units can cluster to pool 128GB of memory for up to 200B-parameter models with up to 1.7x performance gains. [Read more](https://daily.dev/posts/nmAFLDPqB) - **Meta's Muse passes 5 million downloads in 22 days, faster than any prior AI assistant:** Sensor Tower estimates put Muse's download ramp ahead of ChatGPT (56 days), Grok (103 days), and Claude (492 days), helped by Meta allocating it up to half of its daily house ad impressions, while a security researcher found and Meta patched a critical flaw in the same window. [Read more](https://daily.dev/posts/Me2t3HekG) ## Similar posts on daily.dev - [The Plugin Was Trusted. The Code Wasn’t.](https://daily.dev/posts/the-plugin-was-trusted-the-code-wasn-t--ob4senb1p) · Medium · 0 upvotes · 0 comments - [GPT-6.1 Sol in GitHub Copilot](https://daily.dev/posts/gpt-6-1-sol-in-github-copilot-xpbo94oza) · GitHub Changelog · 0 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#security](https://daily.dev/tags/security), [#ai-agents](https://daily.dev/tags/ai-agents), [#github](https://daily.dev/tags/github), [#claude-code](https://daily.dev/tags/claude-code) [View this post on daily.dev](https://daily.dev/posts/gpt-6-1-sol-claims-a-saturated-benchmark-plugin4shell-hits-four-coding-agents-at-once-5x07wvnj2) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/gpt-6-1-sol-claims-a-saturated-benchmark-plugin4shell-hits-four-coding-agents-at-once-5x07wvnj2","headline":"GPT-6.1 Sol claims a saturated benchmark, Plugin4Shell hits four coding agents at once","text":"Today's biggest story is a supply-chain bug called Plugin4Shell that let attackers hijack Claude Code, Codex, Copilot, and Gemini CLI plugins despite SHA pinning, with fixes only partly rolled out. A second CVE shows the same class of risk in DeepSeek-Reasonix Studio's git integration. Meanwhile GPT-6.1 Sol is being claimed to have saturated FrontierMath Tier 4, and a new IEEE-accepted study found cheap LLMs fabricate computation results in 38% of generated methods despite near-perfect code structure. Decision models keep multiplying, with Kev, Strands Decider 2B, and llama.cpp support all landing today.","url":"https://daily.dev/posts/gpt-6-1-sol-claims-a-saturated-benchmark-plugin4shell-hits-four-coding-agents-at-once-5x07wvnj2","datePublished":"2026-10-03T04:17:55.193Z","dateModified":"2026-10-03T04:18:15.638Z","author":{"@type":"Organization","name":"Agentic Digest","logo":"https://media.daily.dev/image/upload/s--V91DY4ls--/f_auto,q_auto/v1772617267/logos/agents_digest","url":"https://daily.dev/sources/agents_digest"},"interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/sources/agents_digest","name":"Agentic Digest"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Agentic Digest","item":"https://daily.dev/sources/agents_digest"},{"@type":"ListItem","position":3,"name":"GPT-6.1 Sol claims a saturated benchmark, Plugin4Shell hits four coding agents at once"}]} ```