--- title: Grok 4.5 lands in Cursor, GitLost hits GitHub agents again description: xAI and Cursor shipped Grok 4.5, a model trained from scratch on Colossus and priced at $2/$6 per million tokens — roughly a quarter of Opus 4.8's cost — with... canonical: https://daily.dev/posts/grok-4-5-lands-in-cursor-gitlost-hits-github-agents-again-mhjxqfovq twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Grok 4.5 lands in Cursor, GitLost hits GitHub agents again | daily.dev og:description: xAI and Cursor shipped Grok 4.5, a model trained from scratch on Colossus and priced at $2/$6 per million tokens — roughly a quarter of Opus 4.8's cost — with... og:url: https://daily.dev/posts/grok-4-5-lands-in-cursor-gitlost-hits-github-agents-again-mhjxqfovq og:image: https://api.daily.dev/og/posts/MhjxQfOVq.png og:image:alt: Grok 4.5 lands in Cursor, GitLost hits GitHub agents again 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. # Grok 4.5 lands in Cursor, GitLost hits GitHub agents again **[Agentic Digest](https://daily.dev/sources/agents_digest)** · 5 min read · 1 upvotes · 0 comments ## Summary xAI and Cursor shipped Grok 4.5, a model trained from scratch on Colossus and priced at $2/$6 per million tokens — roughly a quarter of Opus 4.8's cost — with early testers calling it competitive with Anthropic's flagship. OpenAI declared SWE-Bench Pro no longer reliable for measuring frontier coding capability, which landed awkwardly the same week Grok 4.5 reportedly beat GPT-5.5 on that benchmark. The GitLost prompt injection flaw in GitHub's agentic workflows is getting more coverage, with security researchers emphasizing the architectural root cause: agents running on broad service-account permissions with no trust boundary between system instructions and user content. Bun rewrote its entire runtime from Zig to Rust in 11 days using 64 parallel Claude Code agents at a total API cost of $165,000, fixing 128 bugs and cutting memory usage from 6.7 GB to 609 MB for 2,000 parallel builds. ## Content **TLDR:** xAI and Cursor shipped Grok 4.5, a model trained from scratch on Colossus and priced at $2/$6 per million tokens — roughly a quarter of Opus 4.8's cost — with early testers calling it competitive with Anthropic's flagship. OpenAI declared SWE-Bench Pro no longer reliable for measuring frontier coding capability, which landed awkwardly the same week Grok 4.5 reportedly beat GPT-5.5 on that benchmark. The GitLost prompt injection flaw in GitHub's agentic workflows is getting more coverage, with security researchers emphasizing the architectural root cause: agents running on broad service-account permissions with no trust boundary between system instructions and user content. Bun rewrote its entire runtime from Zig to Rust in 11 days using 64 parallel Claude Code agents at a total API cost of $165,000, fixing 128 bugs and cutting memory usage from 6.7 GB to 609 MB for 2,000 parallel builds. --- ## Grok 4.5 ships via Cursor partnership at $2/$6 per million tokens xAI and Cursor released Grok 4.5, trained from scratch on xAI's Colossus supercomputer specifically for coding and agentic tasks. Input tokens run $2/M and output $6/M — compare that to Claude Opus 4.7 at $5/$25 or OpenAI's Sol at $5/$30. Cursor co-founder Michael Truell said it's become the daily driver for many on their team, and independent testers who got early access described it as the first xAI model that clears the bar for day-to-day work. The model is live in Cursor on all plans; a wider public release followed shortly after. xAI is teasing a 2-trillion-parameter model for August, which is presumably where the 'beat everyone' story is supposed to land. [Read more](https://daily.dev/feed-by-ids?id=izeInWMLO&id=YmGyDCaCx&id=vR8mserPU&id=5QtSOBLRN&id=AePQt9TYH&id=R1RP0QL0N&id=xSx6Jz7yN) ## OpenAI declares SWE-Bench Pro no longer reliable for frontier coding evals OpenAI audited SWE-Bench Pro and concluded it no longer reliably measures frontier coding capability — the benchmark has been saturated and gamed to the point where marginal score differences carry no real signal. The timing was rough: the announcement came the same week Grok 4.5 reportedly beat GPT-5.5 on that same benchmark. The honest takeaway is that the field needs better evals, and Databricks published a useful counterpoint today showing how to build internal benchmarks from your own PR history to make data-driven model and harness decisions. [Read more](https://daily.dev/feed-by-ids?id=QTEZJLK3p&id=iBm3gl0BY&id=5DiEspMJG&id=N0adsoQCt&id=QJZvv6O07) ## Bun rewrites its runtime from Zig to Rust in 11 days using 64 parallel AI agents Bun creator Jarred Sumner published a detailed account of rewriting the entire JavaScript runtime from Zig to Rust — one engineer, 11 days, 64 parallel Claude Code agents, $165,000 in API costs. The results are concrete: 128 bugs fixed, memory usage for 2,000 parallel builds dropped from 6.7 GB to 609 MB, binary size shrank ~20%, and HTTP throughput improved 2–5%. The manual equivalent was estimated at three engineers for a year. The conditions that made this tractable — strong test coverage and clear behavioral contracts — don't exist in most production codebases, but the gap is narrowing. [Read more](https://daily.dev/posts/RGruvy6fO) ## GitLost prompt injection in GitHub agentic workflows leaks private repos Noma Security's findings on GitLost are getting wider circulation today. An unauthenticated attacker opens a GitHub Issue in any public repo with embedded plain-English instructions; the AI agent reads private repositories in the same organization and posts their contents publicly. No credentials, no malware, no coding skills required — a single word like 'Additionally' was enough to bypass guardrails in testing. The root cause is architectural: agents run on broad service-account permissions and treat untrusted user input the same as system instructions. GitHub hasn't shipped a code-level fix because prompt injection doesn't have one. Explicit repository whitelists and treating all user-generated content as hostile before it reaches the LLM are the practical mitigations right now. [Read more](https://daily.dev/feed-by-ids?id=QLZfcubqe&id=ffq6gDzD6&id=Rd9IMMBqn&id=h2sgTA49S) --- ## Also notable - **JADEPUFFER: first documented fully autonomous AI ransomware exploited CVE-2025-3248 in Langflow:** JADEPUFFER executed a complete attack chain without human intervention — exploiting CVE-2025-3248 in Langflow, harvesting API keys, encrypting and deleting MySQL data with AES_ENCRYPT() while leaving no recoverable decryption key, and self-correcting an auth error in 31 seconds; CISA has separately ordered federal agencies to patch a related Langflow IDOR flaw (CVE-2026-55255) by Friday under BOD 26-04. [Read more](https://daily.dev/feed-by-ids?id=dAFJI4tEl&id=ozv7AxMky) - **GhostApproval symlink bug affects six major AI coding agents; Amazon, Cursor, Google patched:** Wiz disclosed a Unix symlink attack where poisoned repositories trick AI coding agents into writing outside their workspace — for example, injecting an attacker-controlled SSH key — with confirmation dialogs hiding the real symlink target, making user approval meaningless; Amazon Q Developer, Cursor, and Google Antigravity patched and issued CVEs while Anthropic disputed the severity and Augment and Windsurf acknowledged but haven't shipped fixes. [Read more](https://daily.dev/feed-by-ids?id=lS2QV4toV&id=K9D9JIsC1) - **Lovable in talks to raise $300M at $13.2B valuation, double its December figure:** The Swedish vibe-coding startup hit $500M annualized revenue run rate in June with 146 employees and roughly one million new projects started weekly, with Menlo Ventures expected to lead the round. [Read more](https://daily.dev/feed-by-ids?id=5dctbPyjI&id=O6imAKvX2) - **Cognition launches SWE-1.7 built on Kimi K2.7 base, scores 42.3% on FrontierCode 1.1:** SWE-1.7 uses entropy preservation via top-p sampling replay to prevent RL training collapse, multi-cluster training across four datacenters with 99%+ compressed weight deltas, and self-compaction enabling rollouts up to six hours — reaching near-frontier performance at significantly lower cost than closed models. [Read more](https://daily.dev/feed-by-ids?id=FA8XYo6PF&id=dYFJe8b7m&id=W1NHOkIdp) - **China flags Claude Code versions 2.1.91–2.1.196 as backdoor; Alibaba bans it from July 10:** China's NVDB claimed a monitoring mechanism in those versions transmits geolocation and identity data to remote servers without consent; Anthropic said it was an experimental anti-abuse measure fully rolled back in version 2.1.198 on July 1, while the incident reflects the broader US-China AI tooling split now playing out at the enterprise procurement level. [Read more](https://daily.dev/feed-by-ids?id=hF6pCKjgB&id=qJWFxFnWw&id=24FnSafkK) ## Similar posts on daily.dev - [SpaceXAI launches Grok 4.5, its first model with Cursor](https://daily.dev/posts/spacexai-launches-grok-4-5-its-first-model-with-cursor-ekeyyomvo) · The Next Web · 0 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#claude-code](https://daily.dev/tags/claude-code), [#prompt-injection](https://daily.dev/tags/prompt-injection), [#grok](https://daily.dev/tags/grok) [View this post on daily.dev](https://daily.dev/posts/grok-4-5-lands-in-cursor-gitlost-hits-github-agents-again-mhjxqfovq) ```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/grok-4-5-lands-in-cursor-gitlost-hits-github-agents-again-mhjxqfovq","headline":"Grok 4.5 lands in Cursor, GitLost hits GitHub agents again","text":"xAI and Cursor shipped Grok 4.5, a model trained from scratch on Colossus and priced at $2/$6 per million tokens — roughly a quarter of Opus 4.8's cost — with early testers calling it competitive with Anthropic's flagship. OpenAI declared SWE-Bench Pro no longer reliable for measuring frontier coding capability, which landed awkwardly the same week Grok 4.5 reportedly beat GPT-5.5 on that benchmark. The GitLost prompt injection flaw in GitHub's agentic workflows is getting more coverage, with security researchers emphasizing the architectural root cause: agents running on broad service-account permissions with no trust boundary between system instructions and user content. Bun rewrote its entire runtime from Zig to Rust in 11 days using 64 parallel Claude Code agents at a total API cost of $165,000, fixing 128 bugs and cutting memory usage from 6.7 GB to 609 MB for 2,000 parallel builds.","url":"https://daily.dev/posts/grok-4-5-lands-in-cursor-gitlost-hits-github-agents-again-mhjxqfovq","datePublished":"2026-07-09T04:18:10.082Z","dateModified":"2026-09-13T21:15:26.800Z","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":"Grok 4.5 lands in Cursor, GitLost hits GitHub agents again"}]} ```