Strategic PartnersIssue 037 records
AWS
A GovCloud-and-guardrails edition. On Oct 30 AWS documented how to run the ChatGPT desktop app and Codex against Bedrock in GovCloud US-West under a locked-down configuration, made OpenAI's Daybreak Red cyber model available to vetted customers in GovCloud US-East, and brought AWS Transform migration and second-generation Outposts racks into GovCloud. Each launch leaves authorization, eligibility or feature gaps with the customer, and none reports a SLED deployment. On the constraint side, an AWS bulletin on the deprecated AgentCore Starter Toolkit requires regenerating affected agents, and an independent preprint shows prompt-caching mistakes, including Bedrock-specific ones, mostly fail silently. Virginia Tech's Amazon-funded ML initiative funds agent reliability research, and WVXU reports the water, power and zoning questions facing a proposed AWS campus ahead of Trenton, Ohio's Nov 3 data center ban vote.
- Evidence records
- 7
- Cross-source patterns
- 2
- Evidence classes
- 5 vendor claim1 independent reporting1 academic research
- Outcomes
- 4 emerging2 cautionary1 mixed
- Source freshness
- 5 new this fortnight2 recent
- Research completed
- 2026-11-02
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
AWS is moving frontier models, agents and migration automation into GovCloud, but every launch leaves authorization, eligibility or feature gaps with the customer
Three Oct 30 AWS releases extend GovCloud: the ChatGPT desktop app and Codex can call Bedrock in us-gov-west-1, OpenAI's Daybreak Red cyber model is available in GovCloud US-East, and AWS Transform migration runs in GovCloud US-West with Outposts racks now connecting to GovCloud. Each comes with a boundary condition. The ChatGPT guide says it does not authorize workloads and turns off many features; Daybreak Red needs OpenAI enrollment; Transform's assessment and modernization stay in commercial Regions. None reports a SLED customer or measured result. The evidence supports more GovCloud options, not faster approvals.
Operating questionFor each new GovCloud AI capability, which Region, features and approvals does the agency actually get, and what still has to be authorized or done outside the boundary?
Supporting evidenceAmazon Web Services (with OpenAI ChatGPT desktop app / Codex)Amazon Web Services; OpenAIAWS Transform's agentic migration tooling arrives in GovCloud (US-West) and second-generation Outposts racks connect to GovCloud, but modernization and assessment stay commercial-only
AWS agent tooling can go wrong without visible errors, so teams need inventory and re-checks after tool or provider changes
AWS's bulletin on the deprecated AgentCore Starter Toolkit says agents imported with affected versions must be re-imported and their deployed artifacts replaced, so upgrading the tool alone does not fix them. An independent preprint finds that most prompt-caching mistakes in agents fail silently, and that a Bedrock cache-lifetime change made correct code defective without a code change. Both point to the same control: track which tool and provider settings produced each agent, and re-test when AWS changes tooling or capabilities. The preprint's sample is open-source projects, not SLED systems.
Operating questionDo you know which tooling versions and provider settings built each production agent, and do you re-test after AWS changes them?
Supporting evidenceAWS security bulletin: deprecated Bedrock AgentCore Starter Toolkit import flaws (CVE-2026-105812, CVE-2026-106032) require re-importing agents, as AgentCore CLI 1.0 becomes the replacementYing Zhang, Ruoqi Guo, Yi Liu, Gelei Deng, Yuekang Li, Yepang Liu, Leo Zhang (authors)
Evidence ledger
Full record · every source keeps its link and limitations
AWS publishes a locked-down configuration for running the ChatGPT desktop app and Codex against Amazon Bedrock in AWS GovCloud (US-West)
An AWS technical walkthrough (Oct 30) says the ChatGPT desktop app, with its Work and Codex components, can send model inference to Amazon Bedrock Mantle in us-gov-west-1 using AWS IAM identities. The guide sets an administrator requirements.toml that forces API-only sign-in, disables browser and computer use, memories, plugins, MCP servers and telemetry, and allows only the Bedrock GovCloud endpoint while denying OpenAI and ChatGPT domains. AWS says processing stays in the selected Region, billing is per token with no seat licenses, and zero data retention is supported. It states plainly that following the guide does not authorize a workload; the agency's own system security plan and authorization process still apply. ChatGPT workspace rules and roles do not govern this setup, so device policy and AWS permissions are the main controls.
Why it matters, evidence and limitations
- Why it matters
- State IT shops and public universities that already hold AWS GovCloud accounts get a documented path to pilot an agentic coding and research assistant without sending prompts to a commercial SaaS tenant. That matters where CJIS, IRS 1075 or state data-residency rules have kept such tools off agency desktops. Relevance is bounded: the guide targets organizations whose security architecture requires GovCloud inference, and most local governments and K-12 districts do not operate GovCloud accounts.
- Evidence and measured results
- Vendor technical how-to with full configuration files, troubleshooting and rollback steps. No customer deployment, usage figure, cost comparison or independent security test is reported.
- Limitations and uncertainty
- Vendor guidance only; no deployment evidence. Requires a desktop app version that supports Bedrock in GovCloud, which AWS tells readers to confirm with OpenAI. Disabling web, browser, plugins and memories limits the tool's usefulness relative to commercial versions. Authorization remains with the customer.
OpenAI's Daybreak Red cyber model reaches Amazon Bedrock in AWS GovCloud (US-East) for eligible, OpenAI-vetted defenders
AWS announced on Oct 30 that Daybreak Red (GPT-5.6 Cyber), part of OpenAI's Daybreak cyber defense initiative, is available on Amazon Bedrock in AWS GovCloud (US-East) and US East (Ohio) for eligible customers. AWS describes it as built for authorized vulnerability research, exploit reproduction and mitigation development, with fewer refusals for that work plus stronger identity verification, monitoring and access controls. AWS says it runs on Bedrock's next-generation inference engine with zero-operator access enforced at the chip, inference data is not used for training, and customers need not share Daybreak Red data with OpenAI. Access requires enrollment in OpenAI's Daybreak programme and then enablement through the AWS account team.
Why it matters, evidence and limitations
- Why it matters
- State security operations centers, fusion centers and large university security teams that run in GovCloud could, if accepted by OpenAI, use a less-restricted model for vulnerability research inside their compliance boundary. Relevance is narrow: eligibility is gated, no SLED customer is named, and most local agencies rely on managed security providers rather than in-house exploit research.
- Evidence and measured results
- One-page vendor availability notice. No benchmarks, eligibility criteria, pricing or customer examples are given.
- Limitations and uncertainty
- Vendor announcement; eligibility criteria and cost are not published. Dual-use capability raises misuse risk if access controls fail. Available only in GovCloud US-East among GovCloud Regions.
AWS Transform's agentic migration tooling arrives in GovCloud (US-West) and second-generation Outposts racks connect to GovCloud, but modernization and assessment stay commercial-only
AWS's public sector blog (Oct 30) bundles three launches. AWS Transform is now available in GovCloud (US-West) and can target both GovCloud Regions, automating discovery, wave planning, landing zone creation, network conversion, server replication and cutover for VMware, bare metal, Hyper-V and database workloads. Only migration capabilities are included; modernization, custom transformation and assessment remain in commercial Regions. Separately, AWS Transform is in scope for FedRAMP Class C (formerly Moderate) in US East (N. Virginia), and AWS Transform MGN is in scope for Class D. Second-generation Outposts racks can now connect to both GovCloud Regions for workloads that must stay on premises. AWS claims Transform can compress transformation timelines from years to months but gives no measured case.
Why it matters, evidence and limitations
- Why it matters
- States with VMware estates facing licensing pressure, and agencies holding CJIS or tax data in GovCloud, can now run automated migration inside that boundary rather than staging work in commercial Regions. Outposts in GovCloud gives a hybrid option for systems such as 911 or courthouse workloads that must remain on site. The split matters for planning: assessment and modernization still need commercial Regions, which some state data policies may not allow.
- Evidence and measured results
- Vendor announcement blog; no customer, timeline or cost data. The years-to-months claim is unsupported in the post.
- Limitations and uncertainty
- Vendor claims only. Capabilities are partial in GovCloud. Outposts availability by country and Region should be checked in AWS FAQs. No SLED customer is cited.
AWS security bulletin: deprecated Bedrock AgentCore Starter Toolkit import flaws (CVE-2026-105812, CVE-2026-106032) require re-importing agents, as AgentCore CLI 1.0 becomes the replacement
AWS bulletin 2026-127-AWS (published Oct 6, marked Important) reports two issues in the open-source bedrock-agentcore-starter-toolkit, versions 0.1.4 through 0.3.13: CVE-2026-105812, code injection allowing arbitrary code execution when a crafted Bedrock Agent is imported and then run or deployed, and CVE-2026-106032, external reference handling that could cause unintended network requests or local file access during import. Version 0.3.14 fixes both. AWS says agents imported with affected versions must be re-imported and their local and deployed artifacts replaced. The toolkit was deprecated on March 27, 2026, and AWS recommends migrating to the AgentCore CLI (@aws/agentcore). On Oct 28 AWS made AgentCore CLI 1.0 generally available with an agentcore migrate command, which makes the migration step concrete now.
Why it matters, evidence and limitations
- Why it matters
- Universities and state agency teams that prototyped Bedrock Agents and moved them to AgentCore with the starter toolkit may hold deployed artifacts produced by a vulnerable importer. The risk is highest where several principals share one AWS account, which is common in campus research accounts and shared state platforms. Why now: the replacement CLI reached 1.0 on Oct 28, so remediation and migration can be scheduled together.
- Evidence and measured results
- Vendor security bulletin with affected versions, fixed version and workaround; acknowledgement to an external researcher. CLI 1.0 availability confirmed from the Oct 28 AWS What's New post. No exploitation data reported.
- Limitations and uncertainty
- Bulletin predates the last run window; included because the Oct 28 CLI 1.0 release makes the recommended migration actionable. No evidence of exploitation. CVE identifiers are as published by AWS.
Amazon-Virginia Tech machine learning initiative enters its fifth year, funding five faculty projects focused on agent reliability, evaluation and efficiency
Virginia Tech announced on Oct 29 that the Amazon-funded Initiative for Efficient and Robust Machine Learning, now in its fifth year and housed in the College of Engineering's Sanghani Center, named a 2026-27 Amazon Fellow and funded five faculty-led projects. Awardees are chosen by a committee of Virginia Tech faculty and Amazon researchers. Projects include SpecRx, diagnosing where tool-using agents first go wrong; a benchmarking and security-testing programme for agentic systems (ASTRA); federated in-context learning for responsible generative AI; reducing redundant computation in multi-agent systems, where early results suggest about 40 percent of wasted computation can be cut; and modular, evolving agent skills. An AWS SageMaker AI director is quoted. Several projects promise open-source benchmarks and an Amazon-facing prototype.
Why it matters, evidence and limitations
- Why it matters
- For research universities, this shows how a hyperscaler research partnership is structured: industry funding, joint selection committees, fellowships with internship access, and deliverables that include vendor-facing prototypes. Research offices negotiating AWS partnerships can compare terms. The agent-evaluation and security-testing work addresses problems public agencies face when assessing agent deployments, though outputs are future deliverables.
- Evidence and measured results
- University news release describing funded projects; one early result (about 40 percent waste reduction) is reported without method or data. No peer-reviewed outputs yet.
- Limitations and uncertainty
- Amazon-funded programme reported by the university; research outcomes are planned, not delivered. The 40 percent figure is an early, unreviewed result. Overlaps with the Research Universities stream; cross-tagged because the research enterprise is the affected SLED segment.
WVXU: Ahead of Trenton, Ohio's Nov 3 vote on banning data centers over 25 MW, AWS details water and power plans for a proposed campus while residents and a township push back
WVXU reports that AWS held an open house on Oct 15 for a proposed 18-building computing facility on about 600 acres near Trenton, a Butler County city of about 14,000. AWS's director of energy and water said the site would use water for cooling about 3% of the year and no more than 21 million gallons annually, supplied by the Southwest Regional Water District, whose plant is rated for 6 million gallons a day and now produces 2 million. AWS said the facility would have about a gigawatt of capacity at peak, would pay its share of grid upgrades with Duke Energy, and would use diesel backup generators, without giving a number. St. Clair Township's zoning commission had just voted to ban data centers; AWS has not bought the land and will set next steps after the election. Trenton council withdrew proposed resident benefits tied to the vote after election-law concerns.
Why it matters, evidence and limitations
- Why it matters
- This is the local-government side of AWS AI infrastructure: a small city, two townships, a regional water district and an electric utility are making decisions that determine where AWS capacity gets built. Why now: Issue 17 is on the Nov 3 ballot, three days after this edition. County and municipal leaders elsewhere face the same questions on water, power cost allocation, generator noise and whether to tie incentives to votes.
- Evidence and measured results
- Independent local public-radio reporting with named AWS, water district and township sources; water figure corrected by the outlet on Oct 16. Company figures are AWS statements, not verified usage.
- Limitations and uncertainty
- Published Oct 15, before the last run; included because the Nov 3 vote is imminent and newly decisive. AWS water and power figures are company claims. Outcome of the vote is unknown at edition time. A separate under-construction AWS facility in Trenton's industrial park is covered in other WVXU reporting.
Preprint: prompt-caching mis-integrations in LLM agents mostly fail silently; a Bedrock cache-lifetime change shows how unchanged code can become defective
Preprint, not peer reviewed (the authors state acceptance at ICSE 2027; no journal reference confirms publication). The authors studied 120 prompt-caching pull requests across 10 popular open-source agent systems spanning Anthropic, OpenAI, Gemini, Bedrock and OpenRouter-style gateways, and found mis-integrations on both the request side and in usage counting, most of which fail silently. They note Bedrock, like Anthropic, requires explicit opt-in and limits cache entries per request, and that Bedrock's opt-in one-hour cache lifetime turned previously correct code into a defect without any code change. One case left Bedrock-routed requests without cache markers. Their CacheLint tool reached 0.71 precision and 0.78 recall on the labelled set and found 32 mis-integrations across 571 projects at 0.86 precision, with 18 upstream fixes.
Why it matters, evidence and limitations
- Why it matters
- Agencies and universities paying per token for Bedrock-based agents, including tools like the GovCloud configurations in this edition, can overpay without any visible error if caching is misconfigured. The finding is independent and directly names Bedrock behaviour, which makes it useful for cost governance when SLED teams build or buy agents on AWS.
- Evidence and measured results
- Empirical study of open-source repositories plus a static analysis tool with precision and recall reported; real-world detection on 571 projects. Not a study of public-sector systems.
- Limitations and uncertainty
- Preprint; sample is open-source agent projects, not SLED systems; Bedrock is one of several providers studied, not the sole subject. Cost impact in dollars is not quantified in the abstract or the passages reviewed.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.