Weekly Digest
Week of 11 May 2026
This week at a glance
This week's digest is anchored by a significant compliance milestone: DTA's Policy for the Responsible Use of AI in Government v2.0 is now mandatory for non-corporate Commonwealth entities, with accountable officials, transparency statements, staff training, and use-case registers all required by mid-2026. Practitioners will also find substantive risk management material across several items, including a documented case of enterprise AI agents being gamed through consumption-based metrics at Amazon, and emerging reports of personal information being surfaced by major commercial chatbots — both carrying direct implications for how agencies govern AI procurement, deployment, and performance measurement. On the international policy front, the week surfaces useful comparative material: NIST is actively refining its Cyber AI Profile with public input, the Alan Turing Institute has published new work on assurance tooling for safety-critical AI and sustainability in defence procurement, and Taiwan's structured national health AI governance model offers a concrete example of separating validation and accountability functions. Australia's exclusion from Anthropic's defensive AI coalition and the biosecurity open letter regarding AI-enabled bioweapons risk are also worth noting for those tracking emerging risk domains at the frontier of AI governance.
Headlines
- AU Gov · AI Policy and Governance Newsletter — May 2026
- Global · Experts Say Divergent Definitions Stall Global AI Governance
- Standards · NIST NCCoE Cyber AI Profile Virtual Working Session Series: Usability of the Profile
- Risk · AI chatbots are giving out people’s real phone numbers
- Tech · Data readiness for agentic AI in financial services
Australian Government1 item
AI Policy and Governance Newsletter — May 2026
Good Ancestors' May 2026 newsletter covers several distinct developments relevant to Australian AI governance. The lead items are: an expert open letter urging Australia to use existing biosecurity powers to screen synthetic nucleic acid imports; questions about whether Australia's AI strategy adequately addresses frontier model training; Australia's exclusion from Anthropic's Project Glasswing defensive coalition as Mythos-class cyber capabilities proliferate; and CAISI's pre-release testing agreements with Google, Microsoft, and xAI. Featured Australian publications include DTA's Policy for the Responsible Use of AI in Government v2.0 (now mandatory), new Federal Court generative AI rules, Finance Department AI procurement rule tightening, and APRA and ASIC cyber-risk letters citing frontier AI as an immediate threat to regulated entities. The EU AI Act is also reported to be softening under industry pressure.
Key points
- Good Ancestors' May 2026 newsletter covers biosecurity-AI risk, Australia's AI strategy, Mythos cyberattack capability, CAISI testing agreements, and DTA policy.
- DTA's Policy for the Responsible Use of AI in Government v2.0 is now mandatory; AI use-case registers due across non-corporate Commonwealth entities by mid-2026.
- Australia is excluded from Anthropic's Project Glasswing defensive coalition; frontier AI cyber risk to critical infrastructure has no current Australian mitigation mechanism.
Implications
- Implement Non-corporate Commonwealth entities should confirm progress toward DTA's Policy for the Responsible Use of AI in Government v2.0 requirements — accountable officials, transparency statements, training, and AI use-case registers — ahead of the mid-2026 deadline.
- Consider Agencies responsible for critical infrastructure or essential government services may want to assess their exposure to Mythos-class cyber threats and whether current bilateral MoU arrangements provide adequate access to AI safety information or defensive tooling.
- Monitor Policy and strategy teams could monitor the biosecurity open letter's outcome, Finance's Commonwealth Procurement Rule updates on AI vendor declarations, and any executive action in the US mandating pre-release AI testing — all with potential implications for Australian frameworks.
Global Regulation & Policy3 items
Experts Say Divergent Definitions Stall Global AI Governance
An opinion piece by Sarosh Nagar and David Eaves (University College London), as reported by Newser and syndicated via Let's Data Science, argues that incompatible definitions of AI - ranging from conversational tools to routine algorithms to hypothetical superintelligence - are blocking meaningful international regulatory coordination. The authors contend that mismatched risk time horizons and the concentration of compute among dominant powers reduce structural incentives for those powers to defer to global governance bodies. For practitioners, the piece reinforces that international regulatory uncertainty is likely to persist until technical standard-setting bodies or state coalitions converge on shared terminology and risk taxonomies. No new rules, agreements, or binding developments are announced.
Key points
- An op-ed by UCL researchers argues definitional divergence is the primary barrier stalling international AI governance.
- Compute concentration among major powers reduces incentives to cede authority to global regulatory bodies.
- This is opinion-based analysis of a known problem - no new agreements, standards, or binding developments are announced.
Implications
- Monitor Policy teams engaged in international AI governance forums may want to monitor whether technical bodies such as ISO/IEC or the ITU publish shared AI taxonomies that begin to resolve definitional fragmentation.
- Consider Agencies developing or updating AI governance frameworks could consider explicitly defining the scope of 'AI' used in their policies to pre-empt the definitional ambiguity described in the analysis.
Taiwan Builds Integrated Health Data Platform for Smart Medicine
Taiwan has launched a national digital health initiative called 'Healthy Taiwan', built around a '3-3-3 Framework' integrating health data spaces, interoperability standards (FHIR), and three dedicated AI governance centres covering responsible AI, external validation, and clinical impact evaluation. The platform connects over 400 hospitals under a Zero Trust architecture, drawing on more than 23 million longitudinal National Health Insurance records. NT$2.94 billion (approximately US$93 million) has been committed to the initiative. The institutional separation of governance, testing, and evaluation mirrors approaches observed in other jurisdictions aiming to reduce conflicts of interest and accelerate regulatory readiness for clinical AI.
Key points
- Taiwan's '3-3-3 Framework' establishes three national AI governance centres for responsible AI, external validation, and clinical impact evaluation.
- The model of separating governance, independent testing, and health-technology assessment may inform Australian digital health AI governance design.
- Source is an opinion piece relayed through a data-science outlet - treat with appropriate caution; limited direct APS applicability.
Implications
- Monitor Agencies with health AI or digital health interests - particularly AIHW, DoHAC, or ADHA - may want to monitor outputs from Taiwan's three AI centres, including any published validation protocols or evaluation frameworks.
- Consider Policy teams developing Australian health AI governance structures could consider whether Taiwan's institutional separation of responsible AI, external validation, and impact evaluation offers a useful structural reference.
ATxSummit Convenes Global Leaders to Shape Asia AI Agenda
ATxSummit 2026, hosted by Singapore's IMDA and Informa Tech, is scheduled for 20-21 May 2026 at Capella Singapore. The summit convenes over 4,000 participants from more than 50 countries, including representatives from the World Bank Group, OECD, NVIDIA, Google, Amazon, and OpenAI, alongside research leaders such as Yoshua Bengio. Plenary themes cover agentic systems, AI for public good, practical AI governance, and workforce evolution. G2G roundtables and workshop sessions may produce model-risk guidelines or procurement frameworks of relevance to APS practitioners, but no session outputs have been published at time of item.
Key points
- ATxSummit 2026 convenes 4,000+ leaders from 50+ countries in Singapore on 20-21 May 2026 to address AI governance.
- Themes include agentic systems, practical AI governance, and AI at national scale - directly relevant to APS practitioner concerns.
- This is an event announcement with no published outputs yet; signal value will emerge from post-summit communiques.
Implications
- Monitor Policy and strategy teams may want to monitor post-summit communiques and G2G roundtable outputs for governance frameworks or procurement models applicable to Australian government AI work.
Standards & Frameworks1 item
NIST NCCoE Cyber AI Profile Virtual Working Session Series: Usability of the Profile
NIST's National Cybersecurity Center of Excellence (NCCoE) is hosting a series of virtual working sessions to gather public input on its Cybersecurity Framework Cyber AI Profile - a tool intended to help organisations manage cybersecurity risks arising from AI adoption. The third session, focused on usability across different stakeholder roles and delivery formats, takes place 12 May 2026. A supporting discussion essay has been released to help participants prepare. The profile is still in draft, with this working series feeding into the next revision.
Key points
- NIST NCCoE is running a virtual working series to refine the Cybersecurity Framework Cyber AI Profile.
- Session 3 focuses on usability across AI roles - users, developers, and deployers - and delivery formats.
- This is a US standards development event; limited direct APS participation value but output worth tracking.
Implications
- Monitor Agencies with cybersecurity and AI governance responsibilities may want to monitor the Cyber AI Profile's development as a potential reference framework for AI-related cyber risk management.
- Consider Policy teams could consider reviewing the preliminary draft and discussion essays to assess alignment with existing Australian Government AI and cybersecurity frameworks.
Risk, Assurance & Ethics10 items
Amazon employees automate tasks with MeshClaw
Multiple news outlets report that Amazon employees used an internal AI agent platform, MeshClaw, to artificially inflate token consumption metrics in response to an 80%-weekly-usage target and internal leaderboards. Employees created agents to automate Slack, email, and code-deploy interactions to hit metrics - a phenomenon internally termed 'tokenmaxxing'. Amazon subsequently restricted leaderboard access and stated usage statistics would not affect performance reviews. The case illustrates two recurring governance pitfalls: raw consumption metrics are poor proxies for value and are readily gamed, and agent frameworks integrated with enterprise tooling introduce operational and security risks when run with broad permissions.
Key points
- Amazon employees gamed internal AI usage metrics by automating token consumption via an agent platform called MeshClaw.
- Illustrates a governance failure: raw consumption metrics as AI adoption KPIs create perverse incentives over genuine productivity gains.
- Security concerns arose from agents running with broad permissions on employee hardware - a least-privilege governance gap.
Implications
- Consider APS agencies developing AI adoption metrics could assess whether their KPIs measure genuine productivity outcomes - such as task success rates or time saved - rather than raw usage proxies like token counts or active-user rates.
- Consider Teams evaluating or deploying AI agent frameworks within government environments may want to consider least-privilege defaults, sandboxing, and audit logging before granting agents broad access to enterprise tooling.
AI chatbots are giving out people’s real phone numbers
MIT Technology Review reports documented cases of AI chatbots—including Google's Gemini—surfacing real personal phone numbers and other PII, likely due to inclusion of such data in model training sets. The privacy data removal firm DeleteMe reports a 400% increase in customer queries about AI-related personal data exposure over seven months, with ChatGPT, Gemini, and Claude most frequently cited. Experts note that the mechanism is poorly understood and there are currently few effective remedies. The pattern covers both accurate PII retrieval and generation of plausible but incorrect contact information attributed to real individuals.
Key points
- AI chatbots including Gemini and ChatGPT are exposing real personal phone numbers drawn from training data.
- DeleteMe reports a 400% rise in customer queries specifically referencing generative AI tools exposing personal data.
- PII leakage from LLMs is directly relevant to APS obligations under the Privacy Act and responsible AI policy.
Implications
- Consider Agencies deploying or evaluating LLM-based tools could assess whether their privacy impact assessments adequately account for PII leakage from training data as a distinct risk vector.
- Monitor OAIC and privacy-focused policy teams may want to monitor how regulators in other jurisdictions respond to this emerging class of AI-driven PII exposure, as it may prompt guidance updates.
New project to build trust in AI for air traffic control
The Alan Turing Institute has announced a project to develop the first open-source toolkit for continuously assessing trust in AI systems used in air traffic control. The initiative targets one of the highest-stakes domains for AI deployment, where reliability and explainability are essential. While the project is UK-based and in early stages, the resulting toolkit and methodologies could be relevant to any jurisdiction grappling with AI assurance in safety-critical infrastructure, including Australian regulators and agencies overseeing aviation or other high-consequence systems.
Key points
- Alan Turing Institute will build the first open-source toolkit for continuous AI trust assessment in air traffic control.
- High-stakes safety-critical AI deployment in aviation offers transferable assurance lessons for Australian regulators.
- No direct Australian mandate or agency involvement - primarily a UK research initiative at this stage.
Implications
- Monitor Agencies with AI assurance responsibilities - including DISR, CASA-adjacent policy teams, or AISI - may want to monitor toolkit outputs as they become available.
Building and procuring sustainable Defence AI will boost force resilience
The Alan Turing Institute has published research arguing that incorporating sustainability considerations into the building and procurement of Defence AI systems strengthens operational resilience. The research is UK-focused and directed at defence procurement contexts. The extracted text is truncated, limiting assessment of specific recommendations. Australian Defence and APS agencies working on AI procurement policy or governance may find the framing of sustainability-as-resilience worth tracking, given parallels in Australian Government architecture and responsible AI procurement principles.
Key points
- Alan Turing Institute research links sustainability measures in Defence AI procurement to increased force resilience.
- Findings are UK-focused but offer transferable framing for Australian Defence AI governance and procurement policy.
- Extracted text is truncated - full substance of the research recommendations is not available from this item.
Implications
- Monitor Defence and procurement policy teams may want to monitor the full Turing Institute report for transferable guidance on sustainability criteria in Defence AI acquisition.
- Consider Agencies developing AI procurement frameworks could consider whether sustainability measures are adequately captured in current APS responsible AI procurement guidance.
Anthropic Declines Chinese Request for Mythos Access
The New York Times reports that a representative from a Chinese think tank privately approached Anthropic officials at a Carnegie Endowment meeting in Singapore to request access to Anthropic's newest model, Mythos. Anthropic declined, and US National Security Council officials were informed and reportedly alarmed. The outreach was characterised as informal rather than an official Chinese government request. The episode illustrates how third-party forums — think tanks, academic convenings, and private meetings — are increasingly scrutinised as potential vectors for frontier AI capability transfer, with API-level access controls and licensing terms becoming the primary governance mechanism.
Key points
- A Chinese think tank representative privately requested access to Anthropic's Mythos model at a Singapore meeting; Anthropic refused.
- US National Security Council officials were alerted and reacted with concern, signalling frontier AI access controls as a live geopolitical issue.
- No technical details about Mythos have been disclosed; the governance significance outweighs the technical content of this report.
Implications
- Monitor Australian agencies involved in international AI engagement or bilateral AI policy discussions may want to monitor whether this episode prompts new US guidance on frontier model access controls or export restrictions.
- Consider Teams advising on Australia's participation in multilateral AI fora could consider how informal access requests at third-party convenings factor into engagement risk assessments.
Establishing AI and data sovereignty in the age of autonomous systems
This sponsored report from EDB (via MIT Technology Review Insights) examines the enterprise movement toward AI and data sovereignty - reducing dependence on centralised cloud AI providers and establishing genuine control over models and data estates. Drawing on a survey of over 2,050 senior executives, the report claims 70% of global executives see a sovereign data and AI platform as necessary for success. NVIDIA CEO Jensen Huang's Davos comments on national AI infrastructure are cited as context. The piece frames sovereignty as both a commercial risk-management issue and an emerging global policy priority, though the vendor origin of the research warrants caution.
Key points
- EDB survey of 2,050+ executives finds 70% believe they need a sovereign data and AI platform to succeed.
- AI and data sovereignty - reducing dependence on centralised cloud AI providers - is increasingly a government and enterprise priority globally.
- This is sponsored content from EDB via MIT Technology Review's custom arm; treat survey figures with appropriate scepticism.
Implications
- Monitor Agencies with cloud AI procurement or data residency responsibilities may want to monitor how AI sovereignty narratives are shaping vendor offerings and policy discourse.
- Consider Policy teams could consider whether the framing of AI and data sovereignty in international discourse aligns with or informs Australian Government positions on sovereign AI capability and cloud data governance.
The shock of seeing your body used in deepfake porn
An MIT Technology Review feature profiles adult content creators affected by AI-generated deepfake pornography, covering non-consensual use of likenesses, financial losses from impersonation scams, and reputational harm. Performers describe contracts signed before AI that now enable retroactive use of their content for model training, and platform takedown processes that fail to keep pace with AI-generated content spread. The piece contextualises AI deepfakes within broader debates about fair use, piracy, and consent - issues that intersect with ongoing litigation in the US and emerging regulatory responses in multiple jurisdictions.
Key points
- Adult performers describe widespread non-consensual deepfake content, financial harm, and reputation damage from AI-generated likenesses.
- Australia's Online Safety Act and proposed mandatory standards for platforms are directly relevant to this harm category.
- Item is a human-interest feature focused on US performers - limited direct APS policy signal beyond existing awareness.
Implications
- Monitor eSafety Commission and OAIC policy teams may want to monitor how US litigation on AI training data consent and deepfake liability develops, as outcomes could inform Australian regulatory approaches.
Decisions Included in Forrester Adaptive Process Orchestration Landscape
Forrester Research has published an Adaptive Process Orchestration (APO) Software Landscape for Q2 2026, covering 35 vendors that combine AI agents with deterministic and nondeterministic automation. The report's vendor criteria highlight governance, auditability, hybrid execution models, and human-in-the-loop decisioning as key requirements when deploying agentic AI in operational workflows. The immediate trigger for this coverage is a PR announcement by Decisions + ProcessMaker on their inclusion in the landscape. For APS practitioners, the more useful signal is Forrester's category definition and evaluation criteria rather than the vendor announcement itself.
Key points
- Forrester's Q2 2026 Adaptive Process Orchestration landscape covers 35 vendors using AI agents in automated workflows.
- Forrester's APO criteria emphasise governance, auditability, and human-in-the-loop controls for nondeterministic AI agents.
- This is a vendor PR announcement about analyst inclusion - limited direct signal for APS procurement decisions.
Implications
- Monitor Procurement and governance teams assessing agentic workflow or process automation tools may want to monitor Forrester's APO criteria as a reference frame for vendor evaluation.
Essay Critiques AI Use Scales' Practical Coherence
An essay published by Stephens Lighthouse, 'Blinded by the (traffic) lights: The intellectual bankruptcy of AI use scales,' critiques institutional frameworks that govern student AI use as increasingly made up of negations rather than enforceable rules. The central argument is that enforcement gaps are treated as someone else's implementation problem, and that the ubiquity of generative AI in everyday tools makes on/off permission models incoherent. The Let's Data Science editorial notes this pattern raises equity concerns and suggests monitoring whether institutions publish concrete enforcement procedures. The core critique - that layered permission frameworks shift accountability to downstream implementers - has some resonance for APS AI policy design, though the essay's primary subject is higher education, not government.
Key points
- An essay argues AI use scales in education are unenforced, incoherent frameworks that absorb critique without delivering accountability.
- The enforcement gap critique has parallels in APS AI policy - layered permission frameworks can diffuse accountability similarly.
- The item focuses on education policy; limited direct applicability to Australian federal agency AI governance contexts.
Implications
- Consider APS policy teams designing AI use frameworks could consider whether their own permission-layering approaches specify clear enforcement owners rather than leaving adjudication to implementers.
Alation launches AI Governance system of record
Alation has announced Alation AI Governance, a commercial product that centralises enterprise AI compliance by registering AI models, agents, and tools in a single inventory, generating model cards, routing approvals through regulation-aware workflows, and producing exportable audit records. The product references the EU AI Act, NIST AI RMF, and ISO 42001, and includes an append-only audit trail and SDK-based ingestion from MLOps platforms. The release is a vendor press release rather than independent evaluation; practical value will depend on connector breadth, evidence provenance, and integration with existing pipelines.
Key points
- Alation has launched a commercial AI governance product providing a centralised inventory, model cards, and audit trail.
- The regulation registry references EU AI Act, NIST AI RMF, and ISO 42001 - frameworks APS agencies already track.
- This is a vendor product announcement; Australian government applicability depends on procurement fit and integration complexity.
Implications
- Monitor Agencies assessing AI governance tooling for model inventory or audit-readiness purposes may want to monitor how this product matures and what independent evaluations find.
- Consider Procurement and AI governance teams could consider whether centralised AI asset registry and model card tooling - commercial or bespoke - is a capability gap in their current governance arrangements.
Technical Developments1 item
Data readiness for agentic AI in financial services
This sponsored content piece, produced by MIT Technology Review's commercial arm on behalf of Elastic, examines data readiness challenges for agentic AI in financial services. It highlights fragmented legacy data, poor indexing, and accuracy demands as key barriers, and positions enterprise search platforms as foundational infrastructure. Use cases discussed - continuous risk monitoring, trade workflow review, and regulatory reporting - are structurally similar to challenges APS agencies face in regulated, high-accuracy environments. The recommended approach of incremental piloting and strong data governance is consistent with APS AI implementation guidance, though the article's commercial framing limits its authority as independent evidence.
Key points
- 57% of financial organisations are still developing internal capabilities to fully leverage agentic AI, per Forrester.
- Agentic AI use cases in regulated sectors - risk monitoring, trade compliance, regulatory reporting - map closely to APS agency contexts.
- This is vendor-sponsored content from Elastic via MIT Technology Review's custom content arm, not independent editorial.
Implications
- Consider APS agencies exploring agentic AI pilots may want to consider whether their own data readiness - particularly legacy data fragmentation and indexing - has been assessed before selecting use cases.
- Monitor Policy and governance teams could monitor how regulated-sector experience with agentic AI (including in financial services) informs emerging APS guidance on agentic AI deployment and oversight.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.