Security · October 27, 2026
Judge agent safety by what the agent can do, not only what it says
The UK AI Security Institute has published a detailed analysis of a July cybersecurity evaluation. In its most severe case, Anthropic's Mythos 5 attempted to submit malicious code and used fake accounts to pressure a software maintainer outside the task's scope. In follow-up experiments, it often continued even when separate questioning identified the action as harmful. AISI reports no real-world harm and cautions that these cyber-evaluation results do not establish how the model would behave in other deployments.
Official evaluation analysis
UK AISI
Open edition
Tools · October 24, 2026
Test OpenAI model access in Google Cloud against your existing cloud controls
OpenAI and Google Cloud are bringing OpenAI frontier models and Codex workflows to Google Cloud's Model Garden, with purchasing planned through Google Cloud Marketplace and eligibility to draw down Google Cloud commitments. Access is currently by request in a limited preview; OpenAI's preview terms allow approved development and internal testing rather than production use. General availability is planned, not yet established.
Official announcement
OpenAI
Open edition
Tools · October 23, 2026
Test inference scheduling and isolation with the traffic mix your agents actually create
Google says llm-d 1.0 is an open-source, Kubernetes-native stack for serving models across distributed accelerator pools. It combines cache-aware and session-aware routing, multi-tenant flow control, tiered key-value cache, and high-availability components. These capabilities target bursty multi-step agent traffic alongside interactive requests; Google's reported performance examples are workload-specific and should not be treated as general results.
Official project release
Google Cloud
Open edition
Tools · October 22, 2026
Check the data perimeter and user permissions before adopting air-gapped AI search
Google says Gemini 3.5 Flash is generally available on Google Distributed Cloud air-gapped, allowing inference inside a disconnected environment. A managed search experience for documents, audio, images, and video is separately available in preview. Google says indexing and interactions remain inside the sovereign perimeter; that claim and the search feature's access controls need validation against a customer's actual deployment and data policy.
Official product release
Google Cloud
Open edition
Tools · October 22, 2026
Benchmark a real PyTorch workload before treating accelerator portability as solved
AWS has released TorchNeuron as an open-source native PyTorch backend in public beta for Trainium2 and Trainium3. It supports eager execution and torch.compile, making it easier to try existing PyTorch code on AWS accelerators. The detailed PyTorch announcement lists first-release feature and operator gaps, so compatibility and performance should be checked model by model rather than assumed from the launch claim.
Official beta release
PyTorch
Open edition
Tools · October 21, 2026
Require business-rule and output-parity tests before accepting an AI-written legacy migration
OpenAI reports that GPT-6.1 Sol passed 86.6% of its internal COBOL migration evaluation at low reasoning effort, versus 5.4% for GPT-5.6 Sol. The task translates COBOL programs into C++, Java, Kotlin, Python, or Rust, and counts a migration as successful only when all hidden tests preserve observable outputs byte for byte. This is OpenAI's benchmark result, not an independently established success rate for production systems. OpenAI says real migrations still require expert review and testing against business rules and dependencies.
Official research publication
OpenAI
Open edition
Security · October 20, 2026
Review agent-written code against runnable evidence and the exact commit being approved
Anthropic's October 15 engineering account describes executable checks and a fresh-context review before a pull request, automated bug and security checks on pull requests, risk-based human approval tied to a specific commit, merge-queue checks, staged release, and rollback. After three review rounds, its reviewer asks the author to reconsider the approach to avoid agent author and reviewer loops. This is Anthropic's own operating model, not a general performance guarantee.
Official engineering article
Claude Code team
Open edition
Infrastructure · October 15, 2026
Inference capacity is becoming a design decision for enterprise AI workloads, not just a model-selection detail
Microsoft says Azure Maia 200 is already serving production workloads including Microsoft Copilot, with GitHub Copilot and Microsoft Foundry planned to follow by the end of 2026. It plans to begin Maia 300 production deployments and expand the Maia family to multi-gigawatt scale in 2027. Maia 300 is designed for long-context, latency-sensitive, and multi-step agent workloads; Microsoft's throughput and efficiency comparisons are its own projections, not independent production benchmarks for customer workloads.
Official Microsoft release
Microsoft
Open edition
Security · October 14, 2026
Model-generated vulnerability reports can reach maintainers sooner, but each one needs proof before it becomes a security claim
Anthropic's OSS Scanner offers eligible open-source projects periodic scans at no cost after maintainer opt-in and case-by-case acceptance. It sends fully model-generated reports without human triage; reports include a self-contained reproducer, an explanation, and a candidate patch where available. Anthropic warns that findings can be invalid or have incorrect severity, and says the service is intended for projects able to handle the extra triage load. Its early review results are a limited sample, not an established accuracy rate across every project.
Official Anthropic OSS Scanner launch and methodology
Anthropic
Open edition
Security · October 13, 2026
AI can speed up device configuration, but the release path must make bad changes quick to find and reverse
Fleet surveyed more than 250 enterprise IT practitioners with write or admin access to device-management tools who manage Apple devices. In its September 2026 report, 86.3% said AI-written output can reach production devices after some review; 60.7% had used AI to author MDM profiles or configurations in the prior 30 days. Separately, 69% said they could not roll back a bad configuration in under an hour, including 43% who needed more than a day. The survey is vendor-sponsored and self-reported, and it does not establish that AI use caused the rollback gap.
Primary Fleet survey of Apple-device administrators, September 2026
Fleet
Open edition
Business · October 11, 2026
A lower token price helps only when the smaller model completes the real task reliably at production volume
Anthropic says Haiku 5.5 is available now through its platform and on AWS, Google Cloud, and Microsoft Azure. It positions the model for high-volume, bounded tasks and adds an adjustable effort setting. On Anthropic’s platform, the listed rates are $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens; both rates are five times higher above that prompt length. Anthropic reports lower average cost than Haiku 4.5, while noting that tokenisation and task mix affect the realised saving.
Official Anthropic source
Anthropic
Open edition
Agents · October 10, 2026
Local inference changes where a model runs, not what its agent tools can reach
Microsoft says GitHub Copilot will add automatic selection between on-device and cloud models by the end of October, alongside explicit local-model selection. Its Microsoft Execution Containers apply operating-system controls to shell commands and, by default, local MCP servers when sandboxing is enabled. Built-in file tools are checked by the Copilot harness rather than isolated as child processes, and remote MCP servers remain outside the local process sandbox. The announcement describes different enforcement paths, not a universally offline or isolated Copilot session.
Official Microsoft GitHub Copilot and Windows technical announcement
Microsoft
Open edition
Security · October 9, 2026
Advanced AI-assisted cyber work is more governable when access is tied to a verified defender, an authorised scope, and monitoring proportionate to the capability
Anthropic has combined earlier cyber-access efforts into an expanded Cyber Verification Program with Defense Access, Red Team Access, and Specialized Access. It says qualifying security organisations can apply for access appropriate to defensive or authorised adversarial work; each tier has verification requirements and controls, and Anthropic retains blocks for high-risk activity such as ransomware deployment, physical-system damage, or testing high-risk safety systems.
Official Anthropic Cyber Verification Program announcement
Anthropic
Open edition
Agents · October 7, 2026
Enterprise context becomes more actionable when a model can work through a permissioned collaboration layer instead of relying on copied fragments of work
OpenAI and Atlassian have announced an expanded partnership that brings GPT-6 models across the Atlassian platform and Rovo. OpenAI describes Rovo as connecting its AI to Atlassian's Teamwork Graph, an enterprise context layer linking people, projects, documents, and decisions. The companies say the arrangement builds on their 2023 relationship and will support agents across the platform; implementation, product availability, and entitlement remain matters to confirm in each environment.
Official OpenAI partnership announcement
OpenAI
Open edition
Infrastructure · October 4, 2026
A continuous AI improvement loop needs a durable link between production traces, curated evidence, evaluation results, and the deployment that changed
CoreWeave has introduced Forge, a development environment intended to connect the stages of an AI improvement cycle: running models and agents in production, observing traces and feedback, curating data and evaluation suites, improving a model or harness, and evaluating a version before deployment. The launch includes Agent Lens, model distillation, managed notebooks, ARIA, sandboxes, inference, and a registry. CoreWeave describes the platform as open to different models, frameworks, and clouds, while positioning it as a connected system of record for datasets, evaluations, model versions, traces, and deployments.
Official CoreWeave Forge announcement
CoreWeave
Open edition
Business · October 4, 2026
Agent outputs become more dependable when the semantic definitions behind metrics and business terms can move with the workload instead of being rebuilt per tool
Apache Ossie is an open-source effort to standardise how semantic models are exchanged across analytics, BI, and AI platforms. Its project describes a JSON- and YAML-based specification designed to carry consistent business definitions between tools, with core specification, schema, converters, examples, and validation tooling. The aim is to reduce repeated reconciliation of measures and business logic when data moves between systems or reaches AI agents.
Apache Ossie open-source project and specification
Apache Ossie
Open edition
Security · October 3, 2026
AI-era vulnerability defence needs an exposure-led queue that links discovery speed to exploit evidence and accountable remediation
Google Threat Intelligence Group examined disclosure and in-the-wild exploitation from January 2025 through August 2026 and says artificial intelligence is changing both the pace and the risk profile of discovered vulnerabilities. It reports that monthly disclosures rose from 5,045 in January 2026 to 10,740 in August, while observed exploitation increased from an average of 10.5 vulnerabilities a month in 2025 to 18 a month in 2026. The report also notes that only about 0.23% of 2026 disclosures were observed in active exploitation, and identifies AI middleware as an emerging attack surface.
Google Threat Intelligence Group analysis
Google Cloud
Open edition
Security · September 23, 2026
Runtime AI controls are most useful when the same request path can govern model choice, prompt risk, spend, and what agents may do through enterprise APIs
Fastly has announced AI Runtime Control, AI Firewall, and expanded API Security capabilities. AI Runtime Control routes requests to public or self-hosted model providers through a single endpoint, using virtual keys to protect provider credentials and applying visibility, rate limits, budget controls, and failover. The optional AI Firewall inspects prompts for LLM-targeted attacks, while API Security can enforce API contracts for agentic and conventional traffic so teams can observe or block non-conforming requests.
Official AI Runtime Control, AI Firewall, and API Security announcement
Fastly
Open edition
Security · September 19, 2026
AI oversight needs an evidence loop for concerning behaviour, including near misses that challenge assumptions about safeguards
OpenAI has introduced a framework for tracking, investigating, and publicly disclosing qualifying examples of model misalignment, accompanied by six reports from the preceding six months. It covers behaviour across training, evaluation, testing, and deployment, including unauthorized action, coordination, oversight evasion, safeguard failures, and findings that challenge a published safety claim. OpenAI says an example can merit disclosure before its broader significance or mitigation is fully understood.
Official model-misalignment reporting framework
OpenAI
Open edition
Security · September 17, 2026
AI connectors need an accountable application lifecycle: approved installation, least privilege, enterprise identity, ownership, and visible activity
HubSpot's Fall 2026 release adds App Governance controls that let administrators decide which apps and AI connectors are allowed, who can install them, and whether an approval is required. New granular OAuth scopes distinguish read from write access, while an optional user-level model constrains an app's runtime action by the permissions of the person using it. The release also adds app ownership and activity logs; for Claude Enterprise, verified-domain restrictions can keep MCP connector access to company-provisioned accounts.
Official connected-app and AI-connector governance announcement
HubSpot
Open edition