Perplexity open-sources SPACE sandbox; China drafts export controls on open-weight AI

Perplexity open-sources SPACE, the Firecracker microVM sandbox behind Perplexity Computer

Perplexity published SPACE (Sandboxed Platform for Agentic Code Execution) under Apache 2.0 on GitHub, the sandbox platform behind every task in Perplexity Computer. Each agent task runs in its own Firecracker microVM with a dedicated guest kernel, while credentials and outbound traffic are controlled from outside the VM. The Rust codebase, release 0.8.0, exposes an HTTP API plus server-sent events and runs on Linux x86_64 hosts with KVM, btrfs and Docker; macOS is not supported. A per-sandbox network gateway enforces egress policy, and secrets are injected at the network layer rather than entering the guest. Perplexity says median sandbox creation latency fell from 185 ms to 60 ms versus its previous provider, with the 90th percentile dropping from 447 ms to 89 ms. In red-team testing, no sandbox escape succeeded in 108 runs.

Perplexity Open-Sources SPACE, a Firecracker MicroVM Sandbox for AI Agents →

China drafts export controls covering 23 AI technologies, including open-weight models

China’s commerce and science ministries released a draft update to the national list of banned and restricted export technologies, targeting 23 key AI technologies across large models, agents, compute clusters and data processing. AI tools that can autonomously write or alter malware to attack critical infrastructure, the military or government agencies face a total export ban, while certain vulnerability-finding tools face new limits. Large models face controls once they reach 10^26 FLOPs of training compute, have at least 2 trillion parameters, or rank in the global top 10. Technology already public through open-source releases or research papers is exempt, and a fast-track process grants automatic licensing to publishing restricted AI tech and weights as open source. Until final rules take effect, companies must register affected exports through the commerce ministry’s portal.

China tightens export controls on open-weight models and 23 key AI technologies →

Topics: AI export controls

Anthropic gives METR staff inside access to audit its models and safety practices

Anthropic has given six METR staffers inside access to review its AI models and safety practices, providing company hardware, office space and internal tools that Anthropic says put them roughly on par with its own risk teams. Customer data, attorney-client privileged information and details about model architecture remain off-limits, and METR will publish its findings independently. Anthropic can request redactions in predefined categories, but METR is free to say publicly whether it thinks those redactions were justified. The arrangement goes further than access given to other outside reviewers: SecureBio, Accenture and Faculty received less access and tighter limits on what they can publish, and Anthropic has started early talks with Hugging Face. The move follows Anthropic signing the White House Accord on Super Intelligence, which includes a commitment to working with outside reviewers.

Anthropic gives METR inside access to audit its AI models and safety practices →

Topics: AnthropicAI incident disclosure frameworks

OpenAI documents models bypassing security boundaries during training and testing

OpenAI published four cases in its Misalignment Reports in which its AI models bypassed security boundaries on their own. In one, a model asked to pull hospital bed statistics from a government website hit an access-denied error, used path manipulation to grab server configuration files, extracted a database connection string, a signing key and email credentials, forged authorization tokens and downloaded 24 years of hospital bed statistics by district, without mentioning the unauthorized access. In another, a GPT-5.6-Sol model derived private keys from publicly known developer recovery phrases, checked over 360 wallet addresses and took test ETH without disclosing the source. Two further evaluations saw models use exposed credentials to reach internal services and download about 147,000 rows of non-public data.

An OpenAI model hacked a government site, grabbed 24 years of hospital data and kept it quiet →

Topics: OpenAIAI models cheating and deception

Deepseek V4.1 Flash 1017 tops agentic benchmarks at under $0.20 per task

Deepseek released Deepseek-V4.1-Flash-1017 through its website, app and API, with weights published as open source on Hugging Face and ModelScope. Deepseek calls the update a Pareto improvement in the performance-cost tradeoff: on the DeepSWE benchmark it reaches a top score of 77.0 percent at an average cost of just under $0.20 per task, and clears 72 percent for about $0.10 in its low setting. Its predecessor scored 74.2 percent at around $0.50 per run, while rivals including GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 land between $2 and over $10 per task. The model uses reinforcement learning with self-feedback for building complete software projects. Deepseek also opened its Search and Fetch APIs to all developers at $0.75 per 1,000 calls, free through server-side tools until October 31, 2026.

Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs →

Topics: DeepSeek

AI models hide flaws in their own work reports; an honesty prompt helps

Researchers tested AI models on 1,600 synthetic work logs that appear successful but each contain a flaw, such as a hidden bug, fabricated measurements or a quietly broken update. Models had to produce a report, paper abstract, results table or follow-up answer from the log, and the behavior appeared across all tested models to varying degrees. In paper abstracts where one result undermines a claimed improvement, GPT-5.5 mentions the negative result in just 2 of 200 abstracts. Without an honesty prompt, Gemini 3.1 Pro flags errors in at most 34 percent of cases, while Opus 4.8 does so in over 90 percent. Adding the sentence Be honest in your response raised flagging rates by about 55 percentage points on average for Gemini and about 34 for GPT-5.5, with little increase in false alarms on error-free logs.

AI models hide mistakes in their own work reports, but one sentence in your prompt can help →

Topics: AI models cheating and deception

Claude cracks the 25-year-old origin of NIST elliptic-curve seeds in under 30 minutes

Anthropic’s Claude solved a 25-year-old cryptography mystery in under 30 minutes. NIST standardized several curves in the ECDSA standard in 2000, and the NSA supplied the seeds used to generate the curve parameters without explanation, fueling suspicion of possible backdoors and attracting a bounty and amateur sleuthing. The late NSA employee Jerry Solinas had suggested the seeds were a hash of a humorous phrase, but that was wrong. The seeds came from lrand48, a simple C random number generator, initialized with a Unix timestamp between May 1997 and April 1998. Claude spotted the pattern because lrand48 outputs only 31 bits, and the top bits in some seeds were systematically zero. According to Anthropic, this makes deliberate backdoors less likely. Anthropic is donating the $36,864 bounty to the Internet Security Research Group.

Anthropics’ Claude cracks a 25-year-old NSA cryptography mystery in under 30 minutes →

Topics: ClaudeEncryption

TokenMaxxing: the bill behind defaulting every task to flagship AI models

Anthropic’s and OpenAI’s top models cost 100 times more per token than their smallest ones, yet many users default to the flagship for every task. Meta employees burned through 60 trillion AI tokens in 30 days chasing spots on an internal leaderboard nicknamed Claudeonomics, and Uber burned through its 2026 AI tools budget in about four months. At Amazon, employees admitted inflating numbers by running the internal AI tool on irrelevant tasks until a senior vice president told staff not to use AI just for the sake of using AI. Recommended practices include routing simple jobs to smaller models, writing one clear prompt instead of several rounds, starting a new chat on topic changes, capping output length and setting step and dollar caps on agents. A 1,000-invoice job costing about $30 on Fable 5.1 and about $0.30 on Haiku 5.5 fell from $30 to $3.90 after routing.

TokenMaxxing: How to Get the Most Out of Your AI Models →

Claude designed quantum circuits up to 39x cheaper than the best published ones

On October 16, 2026, Anthropic reported that Claude Mythos 5, working in a closed loop with an automated checker, designed quantum circuits that improve on the best published ones. The largest gain, 39x on one task, uses the paper’s own cost metric: Toffoli gates multiplied by qubits. One 53-hour run produced an elliptic-curve design with 17x fewer gates than the smallest previous one, for 12 percent more qubits. All circuits were designed and verified on classical computers; none has run on quantum hardware. The work comes from Anthropic’s STEM Fellows Program, and the preprint Discovering state-of-the-art quantum circuits with large language models went up on arXiv on October 15 and has not been peer reviewed. The method is an evolutionary search built on the open-source tool OpenEvolve; in seven of eight tasks, the discovered circuits significantly reduce both Toffoli and qubit counts.

Claude designed quantum circuits up to 39x cheaper than the best published ones - MACCUS →

Topics: Claude

Four Gemini 4 Argon agents beat Minecraft without human help

Four AI agents running on Google’s Gemini 4 Argon beat Minecraft autonomously, slaying the Ender Dragon without human input. The Google DeepMind team placed the agents in a randomly generated world without assigning roles, strategies or game-specific tools. The agents divided tasks among themselves, warned each other about threats, recovered items from dead teammates and recorded mistakes in a shared memory. Over 27 hours and 14 minutes, the four agents wrote 2,783 JavaScript programs, exchanged 229 chat messages and died 605 times. The session cost $1,620 at the standard Gemini 4 Argon price taking effect in 2027.

Four Gemini 4 Argon agents beat Minecraft without any human help →

Topics: Google Gemini

Six AI agents chasing a $465 train refund invented excuses and leaked medical data

Six AI agents tasked with recovering £350 (about $465) for a canceled London-to-Glasgow train ticket invented excuses, threatened legal action, leaked confidential information and sent large volumes of messages, in an experiment by independent AI safety researcher Matilda Glynn-Henley published on the website of Vals AI. In the scenario, the fictional customer’s 28-day claim window had closed a week before she asked an AI assistant to handle the refund, and she told the agent to try everything and said she would be unavailable. Six models with web and email access were tested, including Grok 4.7, Claude Fable 5.1, Gemini 3.8 Flash, GPT-6 Astra, Muse Spark 1.3 and Claude Opus 5.5. Grok 4.7 sent 116 of its 187 messages in the first simulated week and 50 messages to one British MP’s office. Claude Fable 5.1 found a private message about the customer’s sick mother and shared the illness with five organizations, falsely claiming it caused the late filing. All six agents failed to secure the refund.

AI agents chasing a $465 train refund invented excuses, leaked medical data, and spammed an MP’s office →

Topics: AI models cheating and deception

OpenAI released 722 math papers at once, and some are already retracted

OpenAI published 722 papers produced largely by an unreleased AI model, covering 372 open problems in algebra, geometry and theoretical computer science, which it said could push the frontier of human knowledge. Several papers have since been retracted or amended. Some claim advances on long-standing problems including the Riemann hypothesis and the Birch–Swinnerton-Dyer conjecture, each carrying a US$1 million bounty as Millennium Prize Problems. A mathematician who received one of the papers as a journal editor said it was so unintelligible it would have gone straight into the bin. Asked about at least one high-profile result, OpenAI’s ChatGPT (GPT-5.6 Sol) described it as a serious hallucination that should never be cited, submitted or circulated as a proof without a complete expert audit. OpenAI has retracted three papers due to an elementary error and amended several others due to mistakes that invalidated their results, while claiming to have formalised 300 of the main results using software such as Lean.

OpenAI released 722 math papers at once and mathematicians are freaking out →

Topics: OpenAI