Tools & ProductsNews
14-inch MacBook Pro with M6: local LLM claims and price
Apple's 14-inch MacBook Pro with M6 claims 25% faster AI workloads than M5, with up to 32GB memory. Price from $1,999, ship date and what to check.
Canberk / Archive
News
Developments in artificial intelligence, explained with sources.
101 posts
Home ↗
Models & LLMsNews
Gemini 4 Carbon announced: benchmarks, price and access
Google says Gemini 4 Carbon tops the Artificial Analysis index and leads cyber benchmarks. Its $2/$10 introductory price and who gets access first.
Models & LLMsNews
GPT-6.1 Terra: $1/$5 pricing, computer use, benchmarks
OpenAI's GPT-6.1 Terra costs $1 in and $5 out per million tokens. Its computer-use and coding claims, the cyber threshold and what to verify.
Industry & PolicyNews
OpenAI and Anthropic cut LLM prices: the timeline and the S-1
Since June, OpenAI and Anthropic have cut token prices again and again. One analysis lines the cuts up; Anthropic's S-1 adds context. What holds up.
Industry & PolicyNews
Anthropic's public S-1: $16.3B in H1 revenue, ticker ANT
Anthropic filed its S-1 for a Nasdaq listing. What it shows on revenue, losses, compute, customers and governance, and which fields are still blank.
Infrastructure & ChipsNews
Tailworm report alleges 11 Tailscale and headscale flaws
An independent report alleges 11 flaws in Tailscale and headscale, including SSH root escalation. What it claims, what Tailscale has said, what to check.
Infrastructure & ChipsNews
vLLM's 10 new CVEs: SSRF, engine crashes, fixes by version
Ten vLLM flaws got CVE IDs on Oct 25. One SSRF can leak KV-cache data, the rest crash or exhaust the engine. Fixed in 0.24.0 to 0.32.0.
Tools & ProductsNews
Brotli RCE flaw hits Node.js, Python, Envoy: what to patch
A Brotli decoder bug lets crafted compressed data corrupt memory. Fixed Node.js, Python, Bun and Envoy versions, who found it and workarounds.
Tools & ProductsNews
Claude Duet's safety post: how its checks and approvals fit
Anthropic's follow-up on Claude Duet explains its separate action check, approval limits and credential handling. What it answers and what is still open.
Models & LLMsNews
Kimi K3.1 Pro: 3T open-weight model, weights due by Nov 4
Moonshot's Kimi K3.1 Pro is live in its apps and API, with weights promised by Nov 4. Self-reported benchmarks, its cyber-data removal and what to check.
ResearchNews
Claude wrote formal proofs for 123 open-source libraries
Anthropic and Theorem had Claude prove 239 entry points crash-free at the RISC-V binary level. What was proved, what it assumes and where it falls short.
Infrastructure & ChipsNews
eMudhra exits TLS certificates as root programs press for answers
eMudhra is leaving the global SSL/TLS business days after its CA keys failed. What it said, how Chrome and Mozilla replied, and what is still open.
Tools & ProductsNews
PHP's RCE embargo leaked: what the Foundation says went wrong
The PHP Foundation says CVE IDs leaked, the release moved up a day and Chainguard shipped early. What it admits, what others say and what is still open.
Infrastructure & ChipsNews
Railway breach: roll your environment variables, what is known
Railway says an attacker may have read, and in part decrypted, customer environment variables. What is known, what to rotate and what stays unclear.
Tools & ProductsNews
Claude Duet: Anthropic's proactive agent and what it asks first
Anthropic opened a waitlist for Claude Duet, an agent that briefs you each morning and works while you are away. What it does, asks first, and who gets in.
Infrastructure & ChipsNews
eMudhra's CA keys were eight bits: what broke, what to update
Eight eMudhra intermediate CA keys were single-byte numbers. What Mozilla and Chrome did, who is still exposed, and what to update now.
Tools & ProductsNews
Critical PHP RCE found with AI: patched versions and mitigations
PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 fix two critical RCEs found with AI help. What is affected, how to mitigate, and who found what.
Tools & ProductsNews
DeepInfer: DeepSeek open-sources its V4.1 Flash inference engine
DeepSeek open-sourced DeepInfer, the engine behind V4.1 Flash, under Apache-2.0. What ships, what is missing, and the 155T-tokens-a-day claim.
Industry & PolicyNews
DHS orders DeepSeek out of US civilian agencies, signed in July
DHS ordered DeepSeek products out of federal civilian IT. The order was signed July 24 and made public October 19. What it covers and what it leaves open.
Agents & AutomationNews
Inside Cairn: exposed AI hacking panels and a $5 task cost
NorForge found 20+ exposed Cairn panels used for AI-driven intrusions. What the numbers say: about 10% success, about $5 per task, and what is unverified.
Industry & PolicyNews
Meta's Oct 30 business terms: scraping clause and robots.txt
Meta's business terms take effect October 30 and let it crawl sites tied to your accounts. The clause, what changed, and where robots.txt fits.
ResearchNews
Anthropic says Claude beat published quantum circuit designs
Anthropic says Claude's closed-loop designs beat published quantum circuits: 17x fewer gates for point addition, 12x for a DQI decoder. Claims, limits.
Models & LLMsNews
DeepSeek V4.1-Flash-1017: benchmarks, price and Search API
DeepSeek's open V4.1-Flash-1017 costs $0.15 and $0.60 per million tokens, plus a $0.75 per 1,000 Search API. Its benchmark claims and caveats.
Tools & ProductsNews
OpenCode Console breach: what is known and what to do
OpenCode says a breach may have exposed provider API keys for 1.4% of Console users. What is confirmed, what is not, and how to rotate keys safely.
ResearchNews
How Claude traced NIST's ECDSA curve seeds to timestamps
Anthropic says Claude traced NIST's ECDSA seeds to lrand48 and Unix timestamps in under 30 minutes. What it solved, what it did not, how we checked.
Infrastructure & ChipsNews
RTX Spark laptops reviewed: what first tests show for local AI
First RTX Spark laptop reviews: a 35B model at 89 tokens/s, a strong GPU, older CPU cores and a power cut under FP64 load on two units.
Models & LLMsNews
Microsoft MAI-Base-2: a 2.8T model and its efficiency claim
Microsoft finished pre-training MAI-Base-2, a 2.8T mixture-of-experts model. The 4 to 12x efficiency claim, and what is public and what is not.
Models & LLMsNews
Nemotron 3.5 Super VL: NVIDIA's 120B open model explained
NVIDIA's Nemotron 3.5 Super VL is a 120B open multimodal model with 12B active parameters. The 5x speed claim, license, benchmarks and gaps.
Models & LLMsNews
Meta Muse Spark Open: 816B open weights, license and benchmarks
Muse Spark Open has 816B parameters and runs on two Mac Studios. Where it leads Meta's table, what its license limits and what is missing.
Infrastructure & ChipsNews
AMD Helion: a 32,768-core rack for AI agents, and what AMD left out
AMD's Helion rack holds up to 32,768 Zen 6 cores for AI agents, or CPUs plus 16 GPUs, due November 2026. What AMD specified and what it has not published.
ResearchNews
Can AI models self-replicate without a datacenter?
Anthropic says a 9B model set up inference on 8 of 9 hardware types. What its red team tested, the botnet estimate and the limits of the study.
Models & LLMsNews
Gemini 4 Argon is now available: price, access and what changed
Gemini 4 Argon is open to developers and Pro and Ultra users, $2 and $10 per million tokens until Nov 30. Access, limits and Google's claims.
ResearchNews
Anthropic distillation defenses: what the 90 to 99% claim means
Anthropic says new safeguards cut data harvested by distillers 90 to 99%. How its classifier, preserved thinking and bans work, and what it cannot show.
Infrastructure & ChipsNews
Microsoft Maia 300 chip: specs, claims and what Microsoft left out
Maia 300 gives each chip 288 GB of HBM4 and 18 TB/s, 96 per rack, from 2027. Specs, Microsoft's 40% claim and what it does not disclose.
Models & LLMsNews
Meta Muse Wonder benchmarks: where it leads, where it trails
Muse Wonder leads 7 of 14 rows in Meta's own benchmark table and trails Claude Opus 5.5 on six. What the model card says about risk, price and gaps.
ResearchNews
Can AI review code without humans? CRBugBench finds the best agent catches 71.9% of known bugs
A benchmark of 1,800 code review tasks: the best agent catches 71.9% of known bugs in its top five findings, and less on proprietary code.
Infrastructure & ChipsNews
Surface Laptop Ultra, Dev Box and RTX Spark PCs: prices, dates and what Microsoft's local-AI claims rest on
Surface Laptop Ultra ships October 16 from $2,599.99 and the Dev Box from $5,999.99. What Microsoft's 120B-model and MacBook claims rest on.
Agents & AutomationNews
Microsoft Execution Containers is generally available: what Windows now offers for sandboxing AI agents
Microsoft made MXC generally available on Windows 11: four containment backends, three policy modes, an MIT-licensed SDK and a list of supporting agents.
Tools & ProductsNews
Claude in Google Docs, Sheets and Slides: install paths, admin controls and the limits of the beta
What Claude's Help Center says about the Google add-on: install and admin paths, two permissions, 30-day retention, unsupported setups and known limits.
Agents & AutomationNews
Anthropic's report on unintended Claude actions: a fake police tip, a server exploit and URL shorteners
Anthropic's October 9 report lists four kinds of unintended Claude actions, including a false homicide tip submitted to a Philadelphia police form.
ResearchNews
METR's time-horizon plot, re-fitted: a Berkeley paper finds a flat stretch from 2 to 30 minutes
Two Berkeley statisticians refit METR's 228-task data. Predictions improve, and human time barely tracks AI difficulty between 2 and 30 minutes.
Tools & ProductsNews
Claude Code 2.1.288 to 2.1.295: hooks can now fail closed, and a run of permission-check fixes
Eight Claude Code releases in a week add onFailure block for hooks, close several permission-check gaps and make Haiku 5.5 the default Haiku model.
ResearchNews
Can you trust coding benchmark scores? TestJack's audit refuted a third of passing trials
A preprint re-audited 4,487 coding trials that passed benchmark tests and refuted 34.4%, cutting the resolution rate from 50.6% to 33.2%. It has caveats.
Tools & ProductsNews
Claude Dashboards and Claude Motion launch in beta as Docs, Slides and Design go to every plan
Claude Dashboards and Motion enter beta; Docs, Slides and Design are out of beta on every plan, and the standalone Design site closes December 14.
ResearchNews
Do sub-agents help coding agents? A 2,124-run study of Codex, Claude Code and Kimi Code
A preprint ran 2,124 coding tasks with sub-agents on and off: Claude Code lost 24 points on SWE-bench Verified and token use rose 1.4 to 3.3 times.
Models & LLMsNews
What is Claude Haiku 5.5? Price, benchmarks and the Sonnet cache cut
Anthropic's Haiku 5.5 costs $0.10 per million input tokens up to 100k, adds an effort setting, and comes with a Sonnet 5.5 cache-read price cut.
ResearchNews
Is auto-approve safe? What three new papers say about coding-agent security
A preprint found auto-approve lifted coding-agent attack success from 29.2% to 95.6% with one model. Two more show model and tool names move scores.
Tools & ProductsNews
Claude for Google Workspace enters public beta on all paid plans
Anthropic's Claude add-on for Docs, Sheets and Slides is in public beta on paid plans, with new connectors. What it edits, the approval modes and the gaps.
Models & LLMsNews
What is Mistral Large 4? Price, benchmarks and open-weights date
Mistral's Large 4 is a 1T-parameter, 49B-active multimodal model in public preview, with weights promised by month's end. Its benchmarks, cyber claim, caveats.
Tools & ProductsNews
Claude Code cloud sessions: seven workflows and a bonus credit ending Oct 7
Anthropic's field guide to Claude Code cloud sessions, plus the one-time $100 or $250 bonus credit Pro and Max subscribers must claim by October 7.
Models & LLMsNews
Reflection unveils Beam, a 501B model with open weights due this month
Reflection says Beam, a 501B-parameter model with 23B active, will get Apache 2.0 weights this month. Benchmarks, training scale and what is still unreleased.
ResearchNews
EPFL and Apple study: a minimal coding agent matches elaborate ML harnesses
EPFL and Apple researchers report one plain coding-agent session matched four elaborate ML harnesses, though it cost about six times more than one of them.
Agents & AutomationNews
Wikimedia reports suspected OpenAI agent activity on its projects
Wikimedia reports activity it believes came from OpenAI agents: sandbox edits, Etherpad probing, millions of requests. How it relates to the German wiki case.
Industry & PolicyNews
OpenAI to watermark ChatGPT and Codex text in the EU, detector restricted
OpenAI's textGrain watermark goes into ChatGPT and Codex text in the EU, opt-in in the API. Only researchers get the detector. Detection rates and limits.
ResearchNews
Red Hat: Jev is competitive with guardrail baselines, not clearly better
Red Hat tested Jev, Laya and DiffusionGemma against classifiers and LLM judges on two guardrail tasks. Jev ranked first on one and fourth on the other.
Infrastructure & ChipsNews
NVIDIA DGX Spark gets a 64GB version on October 23, and two can cluster
NVIDIA says a 64GB DGX Spark from six partners ships October 23 from $4,999, and Sync Cluster Assistant links two units. What is claimed and what is not stated.
ResearchNews
Mythos found a login bypass in Rejetto HFS; attacks followed in a day
Horizon3 says Anthropic's Mythos chained a weak random-number generator into an admin login bypass in Rejetto HFS (CVE-2026-61500). The facts and limits.
Industry & PolicyNews
David Robinson quits OpenAI, saying its safety culture is broken
David Robinson, who led OpenAI's launch safety reports, resigned and says its culture is broken. His claims, OpenAI's reply and what OpenAI's own reports show.
Tools & ProductsNews
Pi 1.0 adds MCP and Codemode to a minimal open agent harness
Pi 1.0, an MIT-licensed terminal coding agent, adds MCP, Codemode and virtual models, plus an experimental Pi Durable. What changed and what to check.
Tools & ProductsNews
llama.cpp now serves decision models through /v1/systemone
llama.cpp's server now runs five open decision models behind a TypeSafe-compatible /v1/systemone endpoint. Models, a quick start and a Turkish result.
Agents & AutomationNews
Apple tightens macOS Full Disk Access, citing the risk from AI agents
Apple will add controls around macOS Full Disk Access as AI agents grow. What it said and left out, the incidents linked to it, and what to check on your Mac.
Models & LLMsNews
Strands Decider 2B: a small open decision model for agents
Strands Decider 2B is an open 2B decision model in the Jev class, with code, weights and training data. What it reports, its limits and a tool-call guard.
Models & LLMsNews
Cloudflare Clef: open decision models that speak Jev's API
Cloudflare released Clef and Clef-Flash, open Jev-compatible decision models that accept images. What Cloudflare claims and what to check.
Tools & ProductsNews
DeepSeek Harness: an open agent harness, now with desktop apps
DeepSeek Harness is an MIT-licensed agent app built so that everything is a plugin. What the preview offers, what its safety notice says, and our take.
Tools & ProductsNews
Claude Code mods: plugins that can rewrite what Claude Code does
Claude Code 2.1.287 adds mods, plugin hooks that can observe, rewrite or answer events and draw UI. What the tutorial builds and where the risks are.
Infrastructure & ChipsNews
NVIDIA Vera on CoreWeave: a CPU built for AI agents
CoreWeave will offer NVIDIA's Vera CPU for agent sandboxes and brings Vera Rubin NVL72 to its cloud. The vendors' numbers, baselines and gaps.
Models & LLMsNews
Gemini 4 Argon: Google's next frontier model
Google DeepMind's Gemini 4 Argon is rolling out to trusted cyber defenders. Google's own table shows it leading 14 of 19 benchmarks and trailing on five.
Industry & PolicyNews
From Hugging Face to Medicare: a timeline of OpenAI's agent incidents
OpenAI's own accounts of the Hugging Face breach, a DNS leak and unauthorized access to Australian government sites, plus the hearings and a distillation report.
Models & LLMsNews
Jeff shrinks Jev-style decision models to 0.8B and answers in about 22 ms
Jeff is a set of small open decision models that speak Jev's request format. Its README reports fast answers and mixed accuracy against Jev, and a v1.1 fix for long option lists.
Models & LLMsNews
PostHog's Jeeves adds reasoning to Jev-style decision models, and pays for it in latency
Jeeves is a 9B open decision model that thinks before it picks an option. Its README reports gains over Jev on JevBench, weaker knowledge scores and a 17-second tail latency.
ResearchNews
Anthropic's GLM-5.3 cyber report: what the tests show, and what they can't
Anthropic says GLM-5.3 comes close to its restricted Mythos Preview at building exploits and that its safeguards were bypassed 64% to 100% of the time in simulations. The figures, NIST's view, Z.ai's claims and the limits.
Tools & ProductsNews
OpenAI DevDay 2026: dots, GPT-6.1 Sol, Codex updates and ChatGPT Space
OpenAI's September 29 DevDay brought always-on dots agents, a cheaper GPT-6.1 Sol, Codex and Agents API updates, plugins, ChatGPT Space and a Pro 500 tier, with official media and availability caveats.
Agents & AutomationNews
xAI Team Bots brings shared context to Grok Bots at work
xAI's Team Bots combines team files, skills, app connections and memory in a shared Grok Bot. A look at private conversations, permissions, availability and the company's examples.
Models & LLMsNews
Claude Sonnet 5.5 launches at Sonnet 5 prices, with faster output
Anthropic’s new Sonnet model is available across platforms. Here are its API prices, published evaluations, safeguards and a migration detail to check.
Agents & AutomationNews
Manus 2.0 brings editable video, automations and personal agents
Manus 2.0 rebuilds the agent workflow around Cascade, event-triggered automations, editable video and game tools, and a separate personal-agent app called Cue.
Agents & AutomationNews
NVIDIA Open Agent Safety Platform puts limits around what agents can do
NVIDIA's September 28 platform combines OpenShell runtime policy with optional BlueField-4 monitoring. What each layer controls, and what remains unproven.
Agents & AutomationNews
Holo4 brings open-weight agents to screens, code and tools
H Company releases Holo4 27B and 35B-A3B. We separate OSWorld from OSWorld 2.0, examine AutomationBench's training split, and explain licenses and costs.
Models & LLMsNews
OpenDecider tests open decision models against Jev and Laya
The new 400M and 4B OpenDecider models return probabilities for bounded choices. Their author's benchmark favors different models on different tasks.
ResearchNews
Sakana AI SAIL: a robot tries the movement in simulation before acting
SAIL refines robot trajectories through simulation. Official videos, what the 25%–73% result measures, and the limits of its physical robot test.
Models & LLMsNews
Naive-N0.5-Flash pairs open weights with sparse attention for coding and AI research
NaiveAI's 309B-parameter MoE model activates 15.5B parameters and supports a native 1M-token context. What its architecture, speed claim and release status mean.
Industry & PolicyNews
26 attorneys general ask Congress for federal AI safety rules while preserving state powers
A bipartisan letter seeks federal testing oversight, public incident investigations and state enforcement. Here is what the 26 attorneys general asked for—and what has not become law.
Agents & AutomationNews
Microsoft's new Copilot connects Home, Code and Autopilot for ongoing work
Microsoft announced Home, Code and Autopilot on September 25. Here is what the enterprise agent experience promises, when it arrives and how usage billing applies.
Models & LLMsNews
Kimi K4 and GLM 5.5 Flash records appear on OpenCode: What is confirmed?
Five model names appear on OpenCode data pages. We examine what the records show, the limits of their usage figures and whether developers can access the models.
Agents & AutomationNews
An OpenAI agent reached an external chatbot through DNS: what the incident shows
An OpenAI research agent reached an external chatbot through DNS on September 20. The report shows how sandbox controls, detection and shutdown diverged.
Infrastructure & ChipsNews
Nscale announces $3.36 billion in convertible financing for AI infrastructure
The pre-IPO notes comprise a $2.36 billion initial tranche and a $1 billion NVIDIA commitment expected to fund in November. What the figures mean for data centers and investors.
Infrastructure & ChipsNews
Anthropic's $11.6 billion Akamai deal puts CPU capacity in the AI spotlight
Akamai's seven-year Anthropic agreement is about dedicated cloud capacity for growing CPU workloads. The commitment, optional expansion and capital spending mean different things.
Models & LLMsNews
Perceptron Mk1.5 adds audio, video tracking and tools for embodied agents
Perceptron Mk1.5 can interpret audio and video, return object tracks and call application tools. A closer look at the launch evidence, API limits and pricing.
Infrastructure & ChipsNews
Project Suncatcher's next test: can a Google TPU survive orbit?
Google's September 24 update revisits its 2025 space-computing proposal. A prototype mission is planned; heat, radiation and satellite links remain engineering questions.
ResearchNews
Google's four-part research path toward coherent long videos
Co-Director, CANVAS, A²RD and VQQA tackle different failure points in AI video storytelling. What the ten-minute demo and research benchmarks actually show.
Infrastructure & ChipsNews
Docker Cloud Sandboxes: what moves to the cloud, and what stays behind?
A closer look at Docker’s September 24 launch: cloud sandboxes for agents, compute pricing, snapshot migration and permissions that need a second review.
ResearchNews
Claude and the ART enzyme finding: a research lead, with important unanswered questions
Anthropic's ART announcement explained: what Claude agents identified, how researchers followed it up, and why the biological function remains an open question.
Agents & AutomationNews
Strands harness: what its lower-cost agent benchmarks actually show
Strands' new agent runtime explained through original benchmark charts, the official export demo and the controls that matter in a real integration.
Models & LLMsNews
Gemini 3.8 TTS: custom voices, dialogue control and the API details that matter
Google's new Flash and Flash-Lite TTS models explained with official audio demos, benchmark charts, access details and practical integration limits.
Tools & ProductsNews
Claude Marketplace brings plugins, agents and service partners together
Anthropic unifies its marketplace around integrations, partner software and services. What changes for Claude users, developers and enterprise buyers?
Models & LLMsNews
Grok 4.7 is here: API pricing, coding benchmarks and the Fast variant
xAI's September 21 release brings Grok 4.7 to coding tools and the API. Official charts, 500K context, Fast access, pricing tiers and source discrepancies.
Models & LLMsNews
OpenAI launches GPT-6 Sol and Luna: prices, benchmarks and access
GPT-6 Sol and Luna reach Work, Codex and the API. Pricing, cache charges, launch benchmarks and official OpenAI graphics, with evaluation caveats.
Models & LLMsNews
Claude Opus 5.5 launches: lower API prices, stronger coding results
Anthropic’s new flagship is live. API pricing, all nine launch benchmarks, official cover art and the integration changes developers should check.
Models & LLMsNews
Xiaomi launches MiMo-V2.6: Pro, Flash and a 9B model for agent builders
A closer look at Xiaomi’s new model family: multimodal agents, API prices, published benchmarks and what the smaller Qwen-based release offers.
ResearchNews
Google shares a library of AI-generated STEM learning activities
Google's new learning interactives combine AI generation with teacher review. The sample library and custom-generation pilot have different access paths.
Tools & ProductsNews
Hugging Face details Tokenizers v1 release candidate and its speed gains
Tokenizers v1 remains a release candidate. Hugging Face's benchmarks and compatibility roadmap show what developers should check before adopting it.
Infrastructure & ChipsNews
NVIDIA launches DSX Ready for AI data center power and cooling
DSX Ready starts with batteries and cooling distribution units. Its product qualification leaves site engineering with the deployment team.