In one minute
- Western Australia reports 1,380,882 scans, 122 arrests and nine incorrect alerts. The police account is not an independent audit. [1]
- Hong Kong launches Project SAGE. OpenAI will share ChatGPT and Codex allowances from December; xAI announces watermarking by default from 7 November. [3–7]
- CrowdStrike links instructions to agents’ actions. Chinese embodied AI companies raised about $13 billion in the first half of the year. [8][9]
Lead story
AUSTRALIA · BIOMETRIC SURVEILLANCE · PERSONAL DATA
The trial ended on 31 October. Results were announced on 1 November and published by The Guardian on 2 November, Oslo time.
Police compare faces in public spaces with their alert lists
Cameras on marked police vehicles compared passers-by’s faces with images from police databases. Over five months there were 100 deployments at 43 locations, with 1,380,882 scans and 122 arrests, according to police figures reported by The Guardian. These are scans, not necessarily distinct people. Police also reported 146 contacts with registered sex offenders and ten welfare checks. [1]
Commissioner Col Blanch said none of the nine incorrect alerts led to an arrest, although officers approached and questioned the people flagged. Police say images without a match were pixelated and deleted immediately. [1]
The 0.00065% figure does not measure system accuracy
Dividing nine incorrect alerts by all scans gives approximately 0.00065%. That ratio measures the frequency of those alerts across the scanned volume. It neither establishes the proportion of alerts that were correct nor counts wanted people the system missed. It also does not justify claiming 99.99935% accuracy.
Civil Liberties Australia challenges the presentation and calls for scrutiny of the consequences for those affected. Arrests and convictions are different legal outcomes; comparing them is no substitute for a technical assessment of identification errors either. [1]
Withdrawal is followed by an internal evaluation
The vehicles have temporarily left public deployment while police review results, impacts and safeguards. The independent OIC WA warned on 9 July about collective surveillance harms, discrimination and gradual expansion of use. It was not formally consulted and did not contribute to the initial design. Even very brief collection of biometric data carries obligations, the regulator stresses. [1][2]
Why it matters: proportionality requires considering the processing of people who are not wanted, unjustified interventions and avenues for complaints. The arrest count does not answer those questions.
Verification limits: the results come from police statements reported by the newspaper. No complete independent audit or count of unique individuals has been located. European restrictions on biometric identification do not automatically apply in Australia.
Other news
1. Hong Kong launches Project SAGE to examine financial agents’ identity and authority
FINANCIAL GOVERNANCE · 2 NOVEMBER
The HKMA, Hong Kong’s monetary authority, has launched Project SAGE and an expert group to examine AI agents: programs that can carry out tasks on users’ behalf. It also published a discussion paper on banks’ use of these systems. The aim is to develop applications with appropriate supervision and accountability. [3]
According to Reuters, Eddie Yue proposed seeking views on a register of trusted agents and a possible “Know Your Agent” standard. These are proposals, not adopted obligations. The practical question is establishing whom an agent represents, what it is authorised to do and who is responsible for its actions. [4]
Why it matters: identifying the customer does not prove that a program has authority to initiate a transaction. Project SAGE opens that debate; it does not yet establish a binding standard.
2. OpenAI will share ChatGPT and Codex usage allowances from December
AI SERVICES · ANNOUNCED CHANGE
OpenAI’s documentation says Plus, Pro and Business will share a usage allowance across ChatGPT and Codex from December. Usage will be measured in tokens, the units of information processed and generated by a model. The task and model selected will determine consumption; Instant conversations will not draw on the allowance. [5]
Combined remaining usage and reset times will appear in Settings → Usage. On 1 November, BleepingComputer reported an email announcing tests with a small group of users to integrate Work features into ChatGPT. The documentation gives neither allowances per plan nor a specific December date. [5][6]
Why it matters: conversation and coding will share a usage budget. The announcement does not establish that every account will have less capacity. This is a different change from the allowance increases covered on 30 October.
3. xAI announces invisible watermarks for Grok text from 7 November
TRANSPARENCY · AI ACT
xAI plans to enable watermarking by default for seven Grok variants served through its API, including versions 4.7, 4.6, 4.5, 4.3 and three variants of 4.20. The mechanism adjusts token selection to create a statistical signal; it adds no words or hidden characters. It covers global and regional endpoints. [7]
The provider links the measure to Article 50 of the AI Act. 7 November is its technical schedule, not the law’s general application date. The Commission gives 2 August 2026, with a limited transition to 2 December for marking systems placed on the market before August. Opting out requires declaring that a legal exception applies. [7][10]
Why it matters: watermarking can help identify a text’s origin. Detection is statistical, improves with longer text, can be uncertain and requires access by application. It is not a universal AI detector. [7]
4. CrowdStrike links agents’ instructions to what they execute on a computer
CYBERSECURITY · 30 OCTOBER
CrowdStrike describes Agent Graph, Falcon Guardian’s relational layer, as a way to connect users, agents, models and tools with processes, files and network connections. It allows reconstruction of the chain from a request to its effects on a machine, such as a download followed by an installer being run. [8]
For supported agents, integrations can check instructions or tool calls before execution. The provider lists Claude Code, OpenAI Codex, GitHub Copilot, OpenCode and Cursor. These are CrowdStrike’s commercial descriptions of its features; the publication provides no independent assessment of their effectiveness. [8]
Why it matters: saving a conversation does not establish which files an agent opened or which commands it ran. Linking the two helps investigate incidents and attribute actions.
5. Chinese robot makers focus resources on teaching machines new tasks
ROBOTICS · REUTERS REPORT · 2 NOVEMBER
Reuters describes competition shifting from robots’ bodies to the software that lets them learn. According to ITjuzi, Chinese embodied AI companies raised 93.5 billion yuan, about $13 billion, across 322 deals in the first half of 2026: five times the amount a year earlier. This is not money raised today. [9]
X Square, Galaxea, Spirit AI and Robbyant use human demonstrations, sensors and simulation. Spirit AI says it hired about a thousand people to record activities; Robbyant reports 160,000 hours of training material. These are company figures, not audited results. [9]
Why it matters: learning without reprogramming for every task could expand industrial automation. Current trials do not establish general reliability in unfamiliar workplaces or tasks.
Briefs
South Korea seeks parliamentary support for its AI budget. On 2 November, Lee Jae Myung defended a proposal allocating 21.3 trillion won to AI, semiconductors, physical AI and data centres, within an 820.9 trillion won budget for 2027. The figures come from the proposal submitted in September. The news is the appeal to parliament, which must approve it. [11]
HSBC prepares an AI institute in Hong Kong. Georges Elhedery said the bank aims to open it in the first quarter of 2027, with the Hong Kong University of Science and Technology as its anchor academic partner. Reuters reports the announcement; the institute has not opened. [4]
The signal in the noise
Accountability requires reconstructing what happened. An aggregate percentage does not explain an incorrect identification. A customer’s identity does not establish their agent’s authority. A statistical watermark does not certify that content is true, and a conversation log does not prove what was executed.
The common test is evidence of results and their limits. That applies to robotics too: a learning demonstration is not the same as a reliable service.
On the radar
- The lawsuit against OpenAI over medical advice remains unresolved. Bowen Wan’s parents allege that ChatGPT guided her to reduce epilepsy medication before her death. It was filed on 27 October; the 1 November report is follow-up coverage. These are allegations, with no judicial finding of liability. [12][13]
Sources
[1] The Guardian · WA police facial-recognition trial · 02-11-2026, Oslo
[2] OIC WA · Overt Live Facial Recognition Trial · 09-07-2026 · PDF
[3] HKMA · Hong Kong FinTech Week 2026 · 02-11-2026
[4] Reuters / AOL · HSBC institute and Project SAGE · 02-11-2026
[5] OpenAI Help Center · Understanding ChatGPT and Codex usage limits
[6] BleepingComputer · Work features and shared usage · 01-11-2026
[8] CrowdStrike · Agent Graph Shows How Agents Act on the Endpoint · 30-10-2026
[10] European Commission · Transparency obligations under Article 50 of the AI Act
[11] Yonhap · Lee vows bold investment in AI, future growth engines for “growth for all” · 02-11-2026
[12] The Guardian · Parents of woman who died of seizure sue OpenAI · 01-11-2026
[13] Bloomberg Law · OpenAI sued alleging fatal epilepsy medical advice · 27-10-2026