The Threshold ReportAI progress, only when it's real.
Friday, October 9Real: 8 · Security: 6 · Other News: 4 · Upcoming: 0
The Threshold Report
← Threshold Report
Agents+5Oct 8

Image: cloud.google.com

Google launches a Gemini agent for long-running business work

Google launched a Gemini business agent that can work through documents, calendars, code and company data for hours or days. The cloud-based agent connects to business applications and databases, retains memory, writes and runs code, and delegates to subagents (other agents assigned parts of the job). Businesses can choose Gemini or Claude models. Administrators can set sharing permissions, select models and cap spending; coworker agents have their own accounts, email addresses and activity logs.

SOURCES · cloud.google.com · techcrunch.com
Agents+5Oct 9

Image: openai.com

Asana updates StackAI after tests cut model costs 76-fold

Following 144 test runs checking cost, runtime and answer correctness, Asana released changes to its StackAI browser agent today. GPT-6 Astra conducted the study in Codex, mapping the code and building controlled experiments that led to tickets and proposed code changes under human review. Asana expanded browsing-history budgets and prompt caching (reusing previously processed input), and added batch screenshot removal; caching alone cut GPT-6.1 Sol costs fourfold. The optimized Sol workflow completed every tested task correctly at an estimated $0.47 in model costs and about four minutes per run, 76 times cheaper and five times faster than the original production-model workflow.

SOURCES · openai.com
Compute & Infra+5Oct 8

Image: developers.googleblog.com

Google releases ML Drift to speed up on-device AI

ML Drift is available for developers to download and modify, letting them support on-device AI across different graphics chips through one engine. Google released it under Apache 2.0 as both LiteRT's GPU engine (meaning it runs AI work on a device's graphics chip) and a standalone library supporting the OpenGL ES, OpenCL, Metal and WebGPU graphics interfaces. Google is directing developers away from the legacy TFLite GPU delegate (the component that assigns AI work to graphics hardware), which will receive no new features. Google reports up to 40% lower average frame-processing delay for YouTube Shorts effects, up to 30% faster Adobe mobile AI editing and 30% lower model-processing delay for Snapchat effects.

SOURCES · developers.googleblog.com
Science+5Oct 8

Image: The Threshold Report/GPT Image 2.5

HuggingFaceBio releases Carbon-A and 566 million predicted gene regions

Carbon-A is available to download for finding protein-coding regions directly from DNA, including in poorly studied organisms. HuggingFaceBio's model has 1.2 billion parameters (learned settings) and handles major eukaryotic groups (organisms whose cells have nuclei), reducing dependence on species-specific tools and existing RNA data. Its accompanying database contains 566 million predictions across 48,167 assembled genomes from 22,617 biological groups, with coordinates, reconstructed sequences and confidence scores. Carbon-A scored 0.944 on nucleotide F1 (a measure balancing missed and incorrect DNA-letter predictions) across 42 test genomes, beating the team's comparison models; RNA experiments with ActiveSite and UCSD supported some predicted gene structures missing from RefSeq (a reference-sequence database).

SOURCES · huggingface.co · huggingface.co · huggingface.co
Text+5Oct 8

Image: threadreaderapp.com

OpenAI releases a faster GPT-6.1 Sol mode

Yesterday, OpenAI released GPT-6.1 Sol Ultrafast, offering a faster option for coding and app-navigation work. It is available in the API, Codex and ChatGPT Work; Codex and ChatGPT Work access covers Pro 500, eligible usage-based Enterprise and credit-based Edu plans, with Enterprise administrators required to enable it. API prices are $12 per million input tokens and $60 per million output tokens (tokens are pieces of text a model processes), with support for keeping data in the US or EU. OpenAI says Ultrafast approaches Astra's intelligence at up to eight times the speed of Sol Standard.

SOURCES · threadreaderapp.com
Audio+2Video+2Oct 6

Image: github.com

Kandinsky Lab releases downloadable audio-video models with lip-sync

Kandinsky 6.0 combines 44 kHz audio (the sound's sampling rate) with matching video and lip movements, and developers can download the models for local production. Text prompts or starting images produce clips lasting five seconds. The Lite and Pro versions contain 3 billion and 29 billion learned values respectively, and a separate enlargement model raises resolution to 1920×1080. Developers get hardware-specific setup, ready-made generation settings, hosting examples and measured runtimes, plus support for Diffusers, ComfyUI, vLLM-Omni, SGLang and FastVideo.

SOURCES · github.com
Robots+2Oct 7

Image: github.com

NVIDIA releases Long-WAM for robots that use visual history

On Oct. 7, NVIDIA released Long-WAM code and trained models so robotics teams can test systems that use earlier camera observations. The models build on LongLive's video infrastructure, with tools for learning from robot video, training on benchmarks and running evaluations. Long-WAM supports streaming inference (making decisions as observations arrive), with deployment interfaces for YAM, Franka and Unitree G1 robots. Released variants cover manipulation tests, generalist kitchen tasks and training for two-handed robot work, with runtime profiles for RTX 5090, DGX Spark and AGX Thor.

SOURCES · github.com · arxiv.org
Text+2Oct 8

Image: The Threshold Report/GPT Image 2.5

LightOn releases models for reading text, layouts and charts

LightOn released Apache-licensed OCR models (tools for reading text from images) that developers can download to combine document transcription with layout and figure extraction. Models with 0.8 billion, 1 billion and 4 billion parameters (learned values) can optionally return labeled boxes around page elements, image descriptions and chart-data tables. LightOn measured 2.7 seconds to process one page with its 1-billion-parameter model on an H100 graphics chip and published scripts for repeating the evaluations. In LightOn's five-category ParseBench comparison, the 4-billion and 0.8-billion models scored 75.1 and 74.6 overall, ahead of the 35.1-billion-parameter Infinity Parser Pro at 74.3.

SOURCES · huggingface.co
1 of 8

Curious whether AI can do something yet? Submit a question and we’ll track it for free.

Text

Can AI translate a dense legal contract into plain English reliably enough to act on?

Not yet
See the board

Security

Money Moves

Market Watch

The AI stocks that fund managers are adding, trimming, or dropping.

Open the tracker for the latest holdings.

Open Market Watch
AI Entanglements

Coming soon

AgentsOct 9

Image: openai.com

Sophos says AI resolves 52% of managed security cases

Sophos says agents built through OpenAI Daybreak resolve 52% of its Managed Detection and Response cases end-to-end within boundaries calibrated by analysts. They gather evidence, plan investigations and recommend or perform authorized responses, leaving analysts to concentrate on exceptions and consequential decisions. Sophos retains its Notify, Collaborate and Authorise service modes, and potentially destructive actions still require appropriate human oversight. Average response time in cases using the agents fell from about 38 minutes to 89 seconds, described in the case study as a 96% reduction in investigation time.

Positive - Sophos’s agents cut security response times to 89 seconds while keeping actions within analyst-set boundaries and customer authorizations.
SOURCES · openai.com
AgentsOct 8

Image: anthropic.com

Anthropic launches free vulnerability scans and infrastructure defense

Anthropic's Cyber Mission offers a free, opt-in OSS Scanner for open-source software (code people can inspect and modify) and a Critical Infrastructure Defense Program whose founding partners are using Claude to fix vulnerabilities. The scanner sends model-generated reports without human review, including exploit demonstrations, explanations and suggested fixes where available; Anthropic expects more than 90% of reported findings to be real flaws. Engineers and specialist providers help verify and repair vulnerabilities, work that often took months for earlier Glasswing findings and can require waiting until industrial machinery can be changed safely. Projects unable to handle unreviewed results can continue receiving human-verified, coordinated vulnerability disclosures.

Positive - Anthropic’s opt-in scans give open-source maintainers actionable vulnerability reports, while infrastructure partners are already using Claude to close security holes.
SOURCES · anthropic.com · theverge.com
AgentsOct 8

Image: The Threshold Report/GPT Image 2.5

Goodfire releases lower-cost agent monitors through Baseten

Baseten customers can use Goodfire monitors that inspect internal activations (signals inside a model), calling a separate AI reviewer only when a probe flags possible misuse. Customers can choose logging, human review or refusal for offensive hacking, chemical or biological weapons misuse, or reward hacking (gaming the measure used to judge success). Goodfire says four simultaneous probes reuse the model's existing calculations and add less than 2% to the time to first response. In its Kimi K3 tests, the probes caught 93% of malicious hacking sessions and referred 5.5% of harmless sessions for review, at about $185 per million exchanges versus $5,420 for a cheaper model checking every step.

Positive - Goodfire released monitors that flag malicious hacking from internal model signals, making agent oversight cheaper than reviewing every exchange.
SOURCES · techcrunch.com
AgentsOct 7

Image: aws.amazon.com

AWS releases Strands Box to enforce agent actions on Macs

Strands Box lets developers run automatically approved agents on macOS while enforcing action limits outside the agent itself. AWS combines operating-system isolation (restricting what an agent can access) with Dogwood policies (rules that check both attempted tool calls and earlier activity). Its command-line tools, Python code runners and service connectors expose operations for checks that can limit Slack posts, control Git pushes (uploads of code) or cap API calls. AWS says configured rules are deterministic (enforced the same way each time) and cannot be bypassed through prompting.

Positive - AWS’s sandbox combines operating-system isolation with history-aware action limits, constraining agents without relying on them to obey prompts.
SOURCES · aws.amazon.com · theregister.com
Compute & InfraOct 7

Image: theregister.com

PoeLLM botnet compromises more than 3,000 servers, including AI services

The ongoing PoeLLM botnet (a network of infected machines) has compromised more than 3,000 servers, mainly in the US and Western Europe, including many with vulnerable LiteLLM or Ollama services. Black Lotus Labs investigators first encountered it through a compromised Ivanti Sentry server; Lumen disclosed the campaign on Oct. 7. A GitHub-hosted poem encodes changing addresses for the servers directing the botnet. Infected systems mine cryptocurrency while scanning for targets and distributing exploits to spread the botnet.

Negative - The ongoing PoeLLM botnet has compromised more than 3,000 servers, turning infected systems into cryptocurrency miners and infrastructure for further attacks.
SOURCES · theregister.com
TextOct 8

Image: blog.talosintelligence.com

Talos finds malware instructions targeting AI security verdicts

Malware authors embedded instructions aimed at AI analysis systems in 84 samples collected from January 2025 through July 2026, according to Cisco Talos researcher Ryan Fetterman. In matched tests with five local language models, the strongest techniques produced a net shift of about 35% toward verdicts favoring the attacker. More elaborate attempts to scatter instruction templates or intimidate the models often had little effect or backfired. Talos recommends flagging readable embedded instructions as suspicious and separating extracted malware text from the instructions the AI should follow; conventional detection mechanisms are unaffected.

Negative - Instructions embedded in real malware shifted AI verdicts toward benign in tests, giving attackers a way to evade automated security analysis.
SOURCES · blog.talosintelligence.com
1 of 6

Other News

AgentsOct 8

Image: The Threshold Report/GPT Image 2.5

Learned memory lifts a legal agent's full-task pass rate

A GPT-6 Luna legal agent passed every grading criterion on 15.7% of tasks after ten cycles of retaining lessons from reviewed work, up from 2.9%, according to Niko Grupen. The experiment used 196 synthetic legal tasks (artificially constructed cases), split into 110 training and 86 validation tasks, without changing the model's learned values. A Jev-based gate (a check on which lessons could be retained) rejected client-specific or overly narrow notes, and gains transferred to unfamiliar kinds of legal work. Retrieving only relevant lessons halved per-task cost while preserving average grading quality.

SOURCES · x.com
AgentsOct 9

Image: epoch.ai

Epoch tests find gaps in AI agents' research judgment

Across eleven workplace tasks, Epoch AI tested six models on complete deliverables in design, research, coding and data-center analysis. Each model received workplace tools and context comparable to a new hire's, then worked without further intervention. Claude Fable 5.1 and GPT-6 Astra generally led, handling well-defined work reliably while struggling with implicit workplace standards, informative experiment design and recognizing flawed results; the tested open-weight models (available to download and run) also made more basic errors. Epoch largely automates a simpler chart-restyling step after a model met its employee-quality bar.

SOURCES · epoch.ai
AgentsOct 8

Image: blogs.nvidia.com

NVIDIA publishes seven agent-assisted simulation prototypes

NVIDIA published seven agent-assisted prototypes for robots, driving and digital twins yesterday. Developers connected physics, rendering, sensors, assets and code, then tested camera and lidar agreement (lidar measures distance with lasers), collision behavior and tool reach. Astra and Claude Fable 5 agents created two digital twins (virtual versions of physical systems) and improved two others in about three days; a simulated humanoid cleared a hurdle in 64 of 100 trials. In a disassembly experiment, Astra helped design a wrench and successfully remove a suspension component in simulation.

SOURCES · blogs.nvidia.com
ScienceOct 9

Image: The Threshold Report/GPT Image 2.5

Researchers find discrepancies between OpenAI's written proof and code

Researchers at the University of Cambridge and King's College London found at least two discrepancies between OpenAI's written proof and its Lean code (a machine-checkable version of the proof). The result involves a problem derived from the Navier–Stokes equations (equations for fluid motion). A proof checker can validate the code without establishing that it represents the intended mathematical claim. The authors say automatically formalized proofs still need peer review because translating a written argument into code can change what is being proved.

SOURCES · techcrunch.com
1 of 4

Upcoming

Announced or reported, but not yet real or available.

Compute & InfraOct 7

Microsoft and NVIDIA open preorders for local AI PCs

Waiting on: Product availability and software rollout - RTX Spark laptops are scheduled for October 16, 2026, and compact desktops for November 2026; Microsoft plans the Copilot additions within a couple of months of Oct. 7 with Search controls due in fall 2026.

blogs.nvidia.com · arstechnica.com · theverge.com · theverge.com · theverge.com · techcrunch.com
↑ Top