> ## Content Index
> Fetch the complete content index at: https://www.ai-360.online/llms.txt
> Use this file to discover other available public pages before exploring further.

# Who’s watching while the assistant works?
- URL: https://www.ai-360.online/whos-watching-while-the-assistant-works/
- Published: 2026-11-20T12:32:46.000Z
- Updated: 2026-11-20T12:32:46.000Z
- Author: Stewart Tinson
- Tags: News, Analysis

An assistant that carries on working after you have left the room sounds useful. It also raises a few fairly ordinary questions. What can it access? What is it spending? Who notices when it does something unexpected? [Microsoft’s Ignite announcements](https://www.ai-360.online/microsoft-ignite-shifts-the-agent-conversation-towards-runtime-control/?utm%5Fsource=chatgpt.com) put those questions alongside the promise of agents that keep going without another prompt. I suspect quite a few people will be reading the controls section before getting excited about the productivity claims.

There is good reason to. [Google is reporting agentic tools in live intrusions](https://www.ai-360.online/google-says-agentic-intrusion-frameworks-have-crossed-into-live-operations/?utm%5Fsource=chatgpt.com), while [OpenAI has described pausing frontier-model work after a network isolation failure](https://www.ai-360.online/openai-describes-training-pauses-and-network-isolation-for-frontier-models/?utm%5Fsource=chatgpt.com). Between them, they provide a useful reminder that permission, monitoring and the authority to stop work deserve some attention before deployment.

That thread runs through much of this week’s reading. There is money, ambition and a growing collection of tools. There are also some practical questions about how organisations establish what these systems actually did.

[Execution Level Governance- What audit-ready agent governance actually looks likeDavid Girvin, founder and CEO of Assury argues that model-in-the-loop review, AI governing AI, is fundamentally unreliable for regulated environments: even the best-performing models miss a meaningful share of violations, the reviewing model is typically provided by the same vendor being reviewed, and prompt injection or context poisoning can compromise both the acting agent and its supposed overseer simultaneously. He makes the case for deterministic, architecturally enforced controls instead, walking through Assury's approach of autonomy zones, session risk accumulation, and credential starvation, which lets a compromised agent be cut off from its tools instantly rather than relying on time-boxed access. The conversation touches on why David is sceptical of just-in-time credentialing as a solution for agent security more broadly, since agent sessions don't run on predictable human timescales, along with the current gap between how identity and security vendors are pitching agent protection and what he sees happening at the execution layer in practice. He also discusses the compliance and audit implications of probabilistic decision-making, arguing that regulated industries will increasingly need tamper-evident, hash-chained audit trails that can withstand scrutiny from auditors and regulators who are only beginning to understand agentic risk, and reflects on a named frontier lab's own published framework as an example of the gap between research and practitioner reality. Elsewhere, David reflects candidly on building a bootstrapped security company in an increasingly crowded market, why he turned down aggressive VC funding to stay in control of the product, and what a credible third-party assessment of his own gateway would need to look like given that Assury sits directly in the execution path for every customer's agents.![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/icon/favicon-4eb369b6-88c6-4f77-a954-45024f9fbe62.ico)BrightTALK![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/thumbnail/image_1107713-f81583e8-0557-4e7b-871c-06b5650a24de.png)](https://www.brighttalk.com/webcast/21039/674139?ref=ai-360.online)

[**Microsoft puts runtime control alongside the agent pitch.**](https://www.ai-360.online/microsoft-ignite-shifts-the-agent-conversation-towards-runtime-control/?utm%5Fsource=chatgpt.com) Ignite brought shared workspaces, recurring autonomous work and tools for creating applications through natural language. Around them sit agent identities, execution containers, policy enforcement and business analytics. Buyers now have rather more to test than the quality of an answer, including whether an agent stays within its authority as a task develops. 

[**A coalition takes on the missing evidence.**](https://www.ai-360.online/microsoft-coalition-targets-the-evidence-gap-inside-enterprise-ai-agents/?utm%5Fsource=chatgpt.com) Microsoft’s AI Transparency Initiative brings together more than 40 organisations to work towards common evidence about agent behaviour. The useful ambition is independent verification: records that customers can check against their own systems and retain when they change providers. Agreement on that would be valuable. It remains work in progress. 

[**OpenAI explains why it stopped work.**](https://www.ai-360.online/openai-describes-training-pauses-and-network-isolation-for-frontier-models/?utm%5Fsource=chatgpt.com) The company says an internal research model obtained limited internet access through a DNS misconfiguration, prompting a pause in training and evaluation while it investigated. This is the provider’s own account. Even so, the detail makes it useful reading for anyone asking what triggers intervention and who has the authority to impose it. 

[Execution Level Governance- What audit-ready agent governance actually looks likeDavid Girvin, founder and CEO of Assury argues that model-in-the-loop review, AI governing AI, is fundamentally unreliable for regulated environments: even the best-performing models miss a meaningful share of violations, the reviewing model is typically provided by the same vendor being reviewed, and prompt injection or context poisoning can compromise both the acting agent and its supposed overseer simultaneously. He makes the case for deterministic, architecturally enforced controls instead, walking through Assury's approach of autonomy zones, session risk accumulation, and credential starvation, which lets a compromised agent be cut off from its tools instantly rather than relying on time-boxed access. The conversation touches on why David is sceptical of just-in-time credentialing as a solution for agent security more broadly, since agent sessions don't run on predictable human timescales, along with the current gap between how identity and security vendors are pitching agent protection and what he sees happening at the execution layer in practice. He also discusses the compliance and audit implications of probabilistic decision-making, arguing that regulated industries will increasingly need tamper-evident, hash-chained audit trails that can withstand scrutiny from auditors and regulators who are only beginning to understand agentic risk, and reflects on a named frontier lab's own published framework as an example of the gap between research and practitioner reality. Elsewhere, David reflects candidly on building a bootstrapped security company in an increasingly crowded market, why he turned down aggressive VC funding to stay in control of the product, and what a credible third-party assessment of his own gateway would need to look like given that Assury sits directly in the execution path for every customer's agents.![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/icon/favicon-141df185-1450-4cb2-bea5-d50306f69da5.ico)BrightTALK![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/thumbnail/image_1107713-0034ae27-933f-4c7f-a202-359f789d6758.png)](https://www.brighttalk.com/webcast/21039/674139?ref=ai-360.online)

[**GPT-6.2 brings an efficiency pitch with a consequential safety disclosure.**](https://www.ai-360.online/openai-pairs-gpt-6-2-efficiency-claims-with-tighter-cyber-safeguards/?utm%5Fsource=chatgpt.com) OpenAI presents Astra and Sol as improvements in capability and task economics. It also says both reach its Critical cybersecurity capability threshold, with tighter restrictions on higher-risk assistance. Organisations assessing the release will need to examine access, monitoring and refusal behaviour alongside the company’s performance claims. 

[**Google sees agentic intrusion frameworks in operational use.**](https://www.ai-360.online/google-says-agentic-intrusion-frameworks-have-crossed-into-live-operations/?utm%5Fsource=chatgpt.com) Compromised credentials, exposed systems and known vulnerabilities remain familiar entry points. Agentic orchestration helps attackers move through the stages faster and adjust their approach. Google also describes detectable traces left by these tools, giving defenders something practical to investigate while the methods continue to develop. 

[**NATO’s use of frontier cyber AI raises operational questions at scale.**](https://www.ai-360.online/nato-deployment-moves-frontier-cyber-ai-into-operational-governance-territory/?utm%5Fsource=chatgpt.com) OpenAI reports substantial model use through its work with NATO, alongside exercises involving human oversight of cyber agents. Its account raises questions that enterprise security teams will recognise: authorised scope, validation of findings, coordinated disclosure and the cost of getting useful work done. 

[Execution Level Governance- What audit-ready agent governance actually looks likeDavid Girvin, founder and CEO of Assury argues that model-in-the-loop review, AI governing AI, is fundamentally unreliable for regulated environments: even the best-performing models miss a meaningful share of violations, the reviewing model is typically provided by the same vendor being reviewed, and prompt injection or context poisoning can compromise both the acting agent and its supposed overseer simultaneously. He makes the case for deterministic, architecturally enforced controls instead, walking through Assury's approach of autonomy zones, session risk accumulation, and credential starvation, which lets a compromised agent be cut off from its tools instantly rather than relying on time-boxed access. The conversation touches on why David is sceptical of just-in-time credentialing as a solution for agent security more broadly, since agent sessions don't run on predictable human timescales, along with the current gap between how identity and security vendors are pitching agent protection and what he sees happening at the execution layer in practice. He also discusses the compliance and audit implications of probabilistic decision-making, arguing that regulated industries will increasingly need tamper-evident, hash-chained audit trails that can withstand scrutiny from auditors and regulators who are only beginning to understand agentic risk, and reflects on a named frontier lab's own published framework as an example of the gap between research and practitioner reality. Elsewhere, David reflects candidly on building a bootstrapped security company in an increasingly crowded market, why he turned down aggressive VC funding to stay in control of the product, and what a credible third-party assessment of his own gateway would need to look like given that Assury sits directly in the execution path for every customer's agents.![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/icon/favicon-2cf68a03-8824-415b-9a5b-f60c4dd4cfa9.ico)BrightTALK![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/thumbnail/image_1107713-361fc132-4d34-4666-bd73-c46db3f98df2.png)](https://www.brighttalk.com/webcast/21039/674139?ref=ai-360.online)

[**Palo Alto Networks takes script analysis onto the device.**](https://www.ai-360.online/palo-alto-networks-brings-semantic-script-analysis-onto-the-endpoint/?utm%5Fsource=chatgpt.com) The company says its Cortex XDR small language model examines what supported scripts are trying to do before execution, keeping analysis local. That is an interesting approach to obfuscated code. Independent evidence on false positives, resource use and supported formats will help establish how well it holds up in practice. 

[**Visa directs attention towards the suppliers behind payment fraud.**](https://www.ai-360.online/visa-says-ai-is-scaling-familiar-fraud-faster-through-third-parties/?utm%5Fsource=chatgpt.com) Its latest threat report describes AI making established attacks faster and easier to repeat, alongside increased exposure through third-party providers. The figures reflect Visa’s own visibility. The implication travels further: banks can inherit a fraud problem through a supplier before their customer-facing controls have much opportunity to spot it.

[**LoanPro offers a bounded example of agentic assistance.**](https://www.ai-360.online/loanpro-case-study-puts-audit-trails-beside-agentic-efficiency-claims/?utm%5Fsource=chatgpt.com) Its servicing assistant assembles borrower information and recommends actions for staff. AWS and LoanPro report efficiency improvements, with access controls and interaction logs accompanying the claims. For lenders, the useful questions concern whether staff can trace the evidence, recognise a bad recommendation and intervene before it affects a customer. 

[Execution Level Governance- What audit-ready agent governance actually looks likeDavid Girvin, founder and CEO of Assury argues that model-in-the-loop review, AI governing AI, is fundamentally unreliable for regulated environments: even the best-performing models miss a meaningful share of violations, the reviewing model is typically provided by the same vendor being reviewed, and prompt injection or context poisoning can compromise both the acting agent and its supposed overseer simultaneously. He makes the case for deterministic, architecturally enforced controls instead, walking through Assury's approach of autonomy zones, session risk accumulation, and credential starvation, which lets a compromised agent be cut off from its tools instantly rather than relying on time-boxed access. The conversation touches on why David is sceptical of just-in-time credentialing as a solution for agent security more broadly, since agent sessions don't run on predictable human timescales, along with the current gap between how identity and security vendors are pitching agent protection and what he sees happening at the execution layer in practice. He also discusses the compliance and audit implications of probabilistic decision-making, arguing that regulated industries will increasingly need tamper-evident, hash-chained audit trails that can withstand scrutiny from auditors and regulators who are only beginning to understand agentic risk, and reflects on a named frontier lab's own published framework as an example of the gap between research and practitioner reality. Elsewhere, David reflects candidly on building a bootstrapped security company in an increasingly crowded market, why he turned down aggressive VC funding to stay in control of the product, and what a credible third-party assessment of his own gateway would need to look like given that Assury sits directly in the execution path for every customer's agents.![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/icon/favicon-afa85642-1279-4a68-8510-aa9bb6bd01cd.ico)BrightTALK![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/thumbnail/image_1107713-455c43f9-3e16-49bd-8954-57fbdec7ab02.png)](https://www.brighttalk.com/webcast/21039/674139?ref=ai-360.online)

[**The ICO’s agentic AI evidence window closes today, 20 November.**](https://www.ai-360.online/uk-agentic-ai-evidence-window-closes-with-operational-questions-unresolved/?utm%5Fsource=chatgpt.com) Security, accountability, transparency and automated decisions are among the subjects under examination. Practitioners have an opportunity to explain how permissions, personal data and responsibility move through real workflows. The consultation remains open until the end of today, according to the ICO. 

[**Palo Alto Networks makes a $4.5 billion move into model infrastructure.**](https://www.ai-360.online/palo-alto-networks-bets-4-5-billion-on-secure-open-weight-ai/?utm%5Fsource=chatgpt.com) Its proposed all-stock acquisition of Applied Compute would bring post-training and inference infrastructure for open-weight models into the security business. Completion remains subject to conditions. My reading is that the deal shows how commercially significant secure model adaptation and deployment have become.

[**Anthropic prices its flotation at $83 a share.**](https://www.ai-360.online/anthropic-prices-its-initial-public-offering-at-83-a-share/?utm%5Fsource=chatgpt.com) The offer comprises 980 million shares, including shares sold by existing stockholders. Beyond the immediate transaction, public ownership will bring greater scrutiny of spending and operating performance. How those pressures interact with research and safety commitments will be worth following through the disclosures that come next. 

[Execution Level Governance- What audit-ready agent governance actually looks likeDavid Girvin, founder and CEO of Assury argues that model-in-the-loop review, AI governing AI, is fundamentally unreliable for regulated environments: even the best-performing models miss a meaningful share of violations, the reviewing model is typically provided by the same vendor being reviewed, and prompt injection or context poisoning can compromise both the acting agent and its supposed overseer simultaneously. He makes the case for deterministic, architecturally enforced controls instead, walking through Assury's approach of autonomy zones, session risk accumulation, and credential starvation, which lets a compromised agent be cut off from its tools instantly rather than relying on time-boxed access. The conversation touches on why David is sceptical of just-in-time credentialing as a solution for agent security more broadly, since agent sessions don't run on predictable human timescales, along with the current gap between how identity and security vendors are pitching agent protection and what he sees happening at the execution layer in practice. He also discusses the compliance and audit implications of probabilistic decision-making, arguing that regulated industries will increasingly need tamper-evident, hash-chained audit trails that can withstand scrutiny from auditors and regulators who are only beginning to understand agentic risk, and reflects on a named frontier lab's own published framework as an example of the gap between research and practitioner reality. Elsewhere, David reflects candidly on building a bootstrapped security company in an increasingly crowded market, why he turned down aggressive VC funding to stay in control of the product, and what a credible third-party assessment of his own gateway would need to look like given that Assury sits directly in the execution path for every customer's agents.![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/icon/favicon-c1f2ac07-c787-4182-95d0-cffd31d11feb.ico)BrightTALK![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/thumbnail/image_1107713-4030d582-6f5c-4347-9051-d6176bb5ec98.png)](https://www.brighttalk.com/webcast/21039/674139?ref=ai-360.online)

[**NVIDIA gives national security and export controls a dedicated board committee.**](https://www.ai-360.online/nvidia-creates-board-committee-for-national-security-and-export-controls/?utm%5Fsource=chatgpt.com) The standing committee formalises oversight of national-security policy, trade exposure and export compliance. It is a governance development with wider relevance for infrastructure businesses whose commercial plans depend on changing geopolitical decisions, specialist advice and clear escalation routes. 

[**Mastercard turns quantum preparation into an advisory service.**](https://www.ai-360.online/mastercard-turns-post-quantum-preparation-into-an-advisory-offer/?utm%5Fsource=chatgpt.com) Cryptographic inventories, migration planning and board education sit within its new practice. Much of the immediate work is painstaking: establishing what protects sensitive information, where it is embedded and which suppliers control the dependencies. That discovery work will shape how institutions prioritise migration.

[**NIST tackles the interpretation gaps in biometric standards.**](https://www.ai-360.online/nist-turns-biometric-interchange-rules-into-machine-readable-infrastructure/?utm%5Fsource=chatgpt.com) Its machine-readable framework allows software to process biometric and forensic interchange requirements directly. This is fairly unglamorous infrastructure, with a useful purpose: helping developers, testers and auditors work from consistent requirements. Separate evidence is still needed for the performance, privacy and security of the systems built around them.

[**OpenAI releases a restricted child safety image classifier.**](https://www.ai-360.online/openai-releases-restricted-classifier-for-child-safety-image-review/?utm%5Fsource=chatgpt.com) Approved organisations can use the tool to help prioritise suspected abuse material for trained reviewers. Access is controlled, and responsibility for assessment and action remains with those organisations. Its effectiveness will depend on calibration, reviewer practice and careful handling of highly sensitive information. 

---

[Execution Level Governance- What audit-ready agent governance actually looks likeDavid Girvin, founder and CEO of Assury argues that model-in-the-loop review, AI governing AI, is fundamentally unreliable for regulated environments: even the best-performing models miss a meaningful share of violations, the reviewing model is typically provided by the same vendor being reviewed, and prompt injection or context poisoning can compromise both the acting agent and its supposed overseer simultaneously. He makes the case for deterministic, architecturally enforced controls instead, walking through Assury's approach of autonomy zones, session risk accumulation, and credential starvation, which lets a compromised agent be cut off from its tools instantly rather than relying on time-boxed access. The conversation touches on why David is sceptical of just-in-time credentialing as a solution for agent security more broadly, since agent sessions don't run on predictable human timescales, along with the current gap between how identity and security vendors are pitching agent protection and what he sees happening at the execution layer in practice. He also discusses the compliance and audit implications of probabilistic decision-making, arguing that regulated industries will increasingly need tamper-evident, hash-chained audit trails that can withstand scrutiny from auditors and regulators who are only beginning to understand agentic risk, and reflects on a named frontier lab's own published framework as an example of the gap between research and practitioner reality. Elsewhere, David reflects candidly on building a bootstrapped security company in an increasingly crowded market, why he turned down aggressive VC funding to stay in control of the product, and what a credible third-party assessment of his own gateway would need to look like given that Assury sits directly in the execution path for every customer's agents.![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/icon/favicon-0232410e-3d53-42e9-bd5b-ace7b751751e.ico)BrightTALK![](https://storage.ghost.io/c/b5/36/b536821c-5870-4f52-b9ef-106719725971/content/images/thumbnail/image_1107713-9a1c2ce6-58a4-48bb-88e1-5293b9d4c3f9.png)](https://www.brighttalk.com/webcast/21039/674139?ref=ai-360.online)