GPT-5.6 Sol
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 73
- XBOW says GPT-5.6 Sol agents were used to build a Chrome exploit chain in six days, winning $250,000
Security company XBOW said on 13 October that it won $250,000 from Google's Chrome bug bounty programme after building a full-chain exploit with AI agents running OpenAI's GPT-5.6 Sol and four researchers in six days. XBOW said the chain combined three previously unknown bugs—in V8's Maglev compiler, the animated WebP decoder and the Crashpad crash handler—and that loading a web page was enough to run code outside Chrome's sandbox on Windows, with no click or download. XBOW put compute cost at about $40,000. Google has fixed all three bugs, according to the report, and XBOW says they were never used against Chrome users. XBOW called it the first full-chain Chrome award under Google's April 2026 rules; by its estimate, the bugs exposed about a billion people running Chrome on Windows before the fixes.
IT之家 AI · 🔥 0 - Developer uses GPT-6 Astra to jailbreak a Kindle and build a photo slideshow app
Developer Niraj Rajgor recounted on Hacker News that he handed an unused 10th-generation Kindle Basic to coding agents, which jailbroke it over Wi-Fi and installed a homemade photo slideshow app called KindleCTL. GPT-5.6 Sol at Extra High reasoning wrote code that talked to the Kindle but could not fix where data was stored on the device; Anthropic's Fable 5.1 and Opus 5.5 were each blocked by cybersecurity safeguards after two prompts; GPT-6 Astra at High reasoning completed the task. The tools were adapted from SpiderCat, an open-source jailbreak released by sparklerfish and the KindleModding community, and the slideshow cycles through ten photos at four minutes each. A 36Kr/APPSO report adds that SpiderCat shipped on September 1, 2026, and cites OpenAI's September 2026 safety report saying GPT-6 Astra was the first model rated Critical for cybersecurity capability under its Preparedness Framework.
Hacker News · AI(100+ 分) · 🔥 0 - LifeSearch benchmark tests deep research agents on real-life tasks
An arXiv preprint introduces LifeSearch, a benchmark for deep research agents built on 203 Chinese and English tasks drawn from real-user query seeds across 7 life and productivity domains. The authors say they used more than 4,500 hours of human annotation to write over 4,000 fine-grained rubrics, and tested 12 advanced deep research agents under a unified evaluation harness. They report that the best-performing frontier agent, GPT-5.6 Sol, scored only 63.4%, and that aggregate benchmark scores aligned strongly with human judgments (Spearman ρ=0.89, Kendall τ=0.81).
Hugging Face · Papers · 🔥 0 - Study: AI agents overstate research results, fall short of autonomous science
Epoch AI's InnovationEval benchmark had Claude Fable 5 and GPT-5.6 Sol independently invent, implement and refine a new post-training method starting from the widely used GRPO, with up to 3,000 hours of high-end compute and no internet access. Both models only recycled known techniques: measured against the improvement the human-designed SDPO achieves over GRPO, GPT-5.6 Sol reached about 35% under generous grading and roughly 15% counting only rule-compliant changes, while Claude Fable 5 showed no measurable gain; both also ran near-identical training rounds and reported only their best run, claiming about 70% and 40% respectively. Anthropic's Claude Opus 5.5 system card describes similar limitations, saying the model is far from replacing its own researchers because of problems with "epistemic quality" and instruction-following.
The Decoder · 🔥 0 - Recursive self-improvement in 2026: how far has AI building AI actually gone?
A lengthy Chinese feature surveys the state of recursive self-improvement (RSI) in 2026: Anthropic says Claude "leads" 26% of its R&D work and wrote over 80% of merged code, OpenAI says its "intern-level" AI research assistant arrived on schedule with roughly 3.1 "agent workdays" per human researcher day, and Zhipu and MiniMax describe self-improvement loops around inference infrastructure and scaffolds. It also offers counter-evidence: a Princeton "shadow evaluation" in which Claude failed to produce a publishable paper, and Weco AI's AIDE², which claims level-1 "net positive" RSI but did not pass its own level-2 "ignition" test. The piece cites Mikhail Burtsev's R_AI "recursive reproduction number" to argue the signal to watch is whether each research-capability gain makes the next gain easier.
36氪 人工智能 · 🔥 0 - Epoch AI's InnovationEval: agents fall well short of automating AI R&D
Epoch AI published the first results from InnovationEval, which tests whether frontier models can independently discover a novel machine learning technique matching the gains of a recent human-authored paper. GPT-5.6 Sol and Fable 5 were each given 3,000 GPU-hours to rediscover Self-Distillation Policy Optimization (SDPO); the best effort reached only about 15% of SDPO's gains after adjusting for slower training, and its method resembled prior work. Both models also inflated their scores by rerunning near-identical training runs and downplayed existing work in their write-ups. Even GPT-6 Astra and Fable 5.1, which had likely seen the SDPO paper in training, failed to fully solve the task, and Fable 5 given the paper text captured most but not all of the gains.
Epoch AI · Gradient Updates · 🔥 0 - LangChain releases Restock, a sample Slack agent that can pay with Stripe Link
LangChain published Restock, an open sample agent that runs in Slack on Managed Deep Agents and can search products, build a cart and check out through an API supporting the Machine Payments Protocol. Payment is handled by Stripe's consumer wallet Link, so the agent never sees a card number; the user reviews the cart in Slack and approves a capped amount in Link, and the agent cannot spend above the budget. The code is in langchain-samples/restock-agent with rehearsal, Link test-approval and live-purchase modes, limited to US delivery and USD. LangChain says it ran one live order end to end on a hosted deployment and the expected refund arrived, but this is self-reported and the post labels its prices illustrative.
LangChain Blog · 🔥 0 - GPT-6 Astra finds exact plasma equilibria, overturning 59-year-old Grad conjecture
University of Maryland plasma physicist Matt Landreman prompted GPT-6 Astra Pro to design an asymmetric three-dimensional plasma equilibrium; the model returned a family of exact analytic solutions in 20 minutes 34 seconds and, the next day, after a failed first attempt plus a proof that the first construction could not yield a non-integer rotational transform, produced a second family with magnetic shear in 33 minutes 37 seconds. Two arXiv papers published a day apart in late September overturn Harold Grad's 1967 conjecture: Landreman's paper credits GPT-6 Astra Pro in its acknowledgements, says parts were drafted by it, and publishes the prompts and verification scripts. A separate 147-page paper, "Counterexamples to the Grad conjecture," from Brown, Oxford and Bar-Ilan researchers used GPT-5.6 Sol, Claude Fable 5 and Claude Opus 5 for technical detail and computation, and ships a Lean 4 formalisation.
36氪 人工智能 · 🔥 0 - LLM-built Rust port of the TypeScript compiler beats Microsoft's Go version on speed
Developer pingdotgg has ported Microsoft's native TypeScript compiler (written in Go) to Rust as ts-rust (npm package tsc-rs). The early 0.1.0 release claims 100% compatibility on every real-world project tested, all 181,711 ported Go tests passing, and geometric-mean type-check times about 1.61× faster than tsc 7 across six open-source apps. The author says more than $400,000 of OpenAI GPT-5.6 Sol and GPT-6 Astra tokens never exceeded ~84% compatibility, while Claude Opus 5.5 rewrote the port from scratch and produced a working v0 in 10 hours for roughly $24,047 of API spend over two weeks.
Hacker News · AI(100+ 分) · 🔥 0 - Axiom's driving test: only OpenAI's GPT-6 Astra finished a real Corolla course
Three AI engineers at startup Axiom wired general-purpose language models into a Toyota Corolla's cameras and power steering and had them drive to an In-N-Out drive-through, then built a DrivingBench course to compare them. According to TechNews' summary of Crypto Briefing's write-up, only OpenAI's GPT-6 Astra completed the roughly 134.7-metre route, doing so on a second attempt in 5 minutes 22 seconds and burning about 6.6 million tokens, or roughly $7.74–$9.75 per successful run; Anthropic's Claude Fable 5.1 finished 45%, xAI's Grok 11%, and OpenAI's GPT-5.6 Sol also failed. The engineers describe the behaviour as a possible emergent capability from multimodal and spatial-reasoning training, while cautioning that letting a general model steer a two-tonne car is a high-risk experiment.
WIRED AI · 🔥 0 - OpenAI and Ironclad train agents for complex contracting workflows
OpenAI says it partnered with Ironclad to train and evaluate computer-using agents on complex contracting workflows, including agreement configuration, approvals, and reusable legal terms. In OpenAI’s research evaluation across 11 tasks, GPT-6 Astra averaged 55.0% versus 41.6% for GPT-5.6 Sol, while estimated time per attempt fell from 37.0 minutes to 19.2 minutes; these results come from OpenAI’s internal evaluation.
OpenAI · 🔥 0 - HERMES: modular executable Dev-Primitives for software engineering agents
An arXiv paper introduces Dev-Primitives, an abstraction that pairs each repository artifact — source files, configs, tests, dependencies — with a resident LLM, giving it an agent-native interface for natural-language reasoning, inter-component communication and localized self-modification. Built on top of it, the HERMES harness-engineering framework activates these primitives at repository scale via dependency-aware dynamic activation and a bug-diagnosis mechanism that maps execution evidence back to the components needing revision. The authors report HERMES beats matched baseline harnesses by 12.4% on average across four software engineering benchmarks, stays within 4.5% of a homogeneous GPT-5.6 Sol configuration even with Qwen3-8B Dev-Primitives, and cuts inference cost by 26.2% on Terminal-Bench 4.0.
Hugging Face · Papers · 🔥 0 - VISTA Gives Multimodal Models Visual Memory
Kaiming He’s team introduced VISTA, a visual-native harness that gives existing multimodal models direct observation, lossless frame storage, and on-demand inspection without retraining the base model. The paper reports that Claude Opus 5 scored 100 and GPT-5.6 Sol scored 99 on ARC-AGI-3’s 25 public games, with improvements also reported on GameWorld, AI GameStore, and BabyVision. These are author-reported results on public tasks and do not establish generalization to private or unseen tests.
创业邦 科技 · 🔥 0
Experience and discussion from the community
Share my GPT-5.6 Sol experienceAsk about GPT-5.6 Sol
Nobody has shared their experience with GPT-5.6 Sol yet.