dhx

The use-after-free occurs within ip_options_build (net/ipv4/ip_options.c) which CPU #2 processing the retransmitted SYN packet uses to read/copy options out of the unallocated/reallocated memory that CPU #1 had already freed up. Remember that the options header is variable length too so CPU #2 is seemingly under-reading/over-reading this variable length options header by some fairly random amount between 1-256 bytes depending on what data has been written to the reallocated memory.[1]

The undocumented part (per LKML discussions at least) is what happens next in the attack. Seemingly CPU #2 or CPU #3[2] comes along later to process the random jibberish options header and rarely (in Anthropic's case it took 48min locally) that random jibberish will be returned to the attacker (e.g. copy jibberish options header to a response packet?) and it'll expose a return pointer to a fixed kernel function. The attacker then crafts a SYN packet and retransmits it with an options header that places a return pointer in the right place, and hopes for another race to occur where CPU #3[2] executes a callback to an attacker-controlled pointer stored within the options header data. Anthrophic's case said it took 5min for this stage of the attack to work locally.

Someone will probably write this all up at some point and it'll be quite a long and detailed read of how the 48min and 5min stages of the attack work and perhaps this is where MPTCP code paths are involved due to the way they process the TCP options header data.

[1] https://en.wikipedia.org/wiki/Internet_Protocol_Options

[1] Author of the kernel patch suggests that CPU #3 (three way race) is needed per https://lore.kernel.org/netdev/CAL4WiiroQ8XyDzUnmuy5FmJ4SWfZ...

For more on why MPTCP is not relevant--my limited understanding is the issue is with SYN retransmits.

SYN packet #1 is received and added to a request queue for processing.

SYN packet #2 is received and is a retransmit of SYN packet #1. It is matched to SYN packet #1 already in the request queue and now there is shared state between both SYN packets #1 and #2 in the request queue.

CPU #1 takes on the task of processing SYN packet #1.

CPU #2 takes on the task of processing SYN packet #2.

CPU #1 responds back with RST packet #1 (e.g. requested TCP port is not open) and _should_ defer the destruction of the request socket for SYN packet #1, as other CPUs may still be processing the shared state for the request socket as is the case for retransmitted SYN packets.[1] However, this deferred destruction wasn't occurring, and CPU #1 was _immediately_ destroying the request socket/shared state for SYN packet #1.

CPU #2 races with CPU #1 and reads from the request socket/shared state for SYN packet #1 after its _immediate_ destruction by CPU #1.

Linux's IPv6 handling is unaffected because in this IPv6 scenario, there is no shared state between SYN packets #1 and #2 in the request queue. Instead, each request socket has duplicate data.

Someone more knowledgeable can confirm/correct me on the above please.

[1] https://www.kernel.org/doc/html/latest/RCU/whatisRCU.html#rc...

Fixed upstream in:[1]

7.3 (2026-10-18)

7.2.10 (2026-10-11)

6.18.56 (2026-10-11)

6.12.113 (2026-10-15)

6.6.159 (2026-10-15)

6.1.190 (2026-10-15)

5.15.225 (2026-10-15)

[1] https://lore.kernel.org/all/?q=tcp_v4_reqsk_destructor -- check for "kfree_rcu" being used instead of "kfree" within "tcp_v4_reqsk_destructor" (net/ipv4/tcp_ipv4.c)

See order 5(h) that sets out the limited conditions by which Apple can automatically remove, disable or block an application that has already been distributed and installed on a device. Importantly for Australia this list includes 5(h)(iii) for section 128 of the Online Safety Act 2021 and the receipt of an "app removal notice" from the Australian Government.[1] You can ignore the "subject to orders 12 and 13" as that just addresses timing of the orders coming into effect. An order in this judgement is anything numbered under the heading "AND THE COURT ORDERS THAT:" and preceding the heading "REASONS FOR JUDGMENT".

The nuance is the Online Safety Act 2021 does not have an "app distribution blocking notice" or "app blacklisting notice" or similar, therefore, I think the effect is Apple can't stop a Russian website distributing apps to Australian users who seek them out, but Apple can use a received "app removal notice" to then blacklist the relevant app on existing devices?

[1] https://www.legislation.gov.au/C2021A00076/latest/text

I think it's the wrong question to ask in 2026. By the time a CVE is published and then a new version of software released it's too late--the vulnerability is very likely to have been used in the wild for some period of time already, quite possibly by multiple actors who have independently discovered the vulnerability. Patching is essentially now an ineffective control for mitigating security risks--other controls have to be used instead.

Examples of the type of alternate questions to ask in 2026 may include:

- Does my software/system have a robust architecture where a failure of one component (e.g. RCE in a PHP process) has a minor impact, rather than a catastrophic impact? For example, a minor impact might be just the opportunity for an attacker to chain a PHP RCE with kernel LPE for some common syscall, rather than mass loss of every user's data to the attacker. I raised some PHP web application specific ideas in another thread at [1] but the answer to this question is product-agnostic and should consider microservices architecture vs. monolithic web application with access to everything, and use case/requirements development that considers matters such as users wanting to isolate their own access. For example, allowing bank users read only login for bank accounts vs. separate login for making transfers. Or implementing business process changes such as delayed commit for sensitive user actions--e.g. force delay a bank transfer by 24h and notify the user via various means giving them a chance to undo the transfer before they no longer can.

- How can I keep up-to-date with the latest software/system security architecture trends? What is the equivalent of [2] for learning about architecture patterns to potentially use for a simple eCommerce site, HR/finance web application, etc in 2026? For components of a web application such as interpreter, database server, etc--what options exist for each and how good is the project's choice of software languages, dependencies, etc to offer the strongest security protections in 2026? When was the last time I checked if someone had written an interpreter for PHP, Python, etc in Rust, SPARK, or whatever, or written an interpreter with much more comprehensive unit testing included?

- How quickly can my users or I detect security incidents / deviations of system behaviour from typical state? For example, [3] was discussed on HN today as once such detection mechanism where canary tokens are deliberately littered throughout a system or dataset. The security incident may have already happened, but early detection forces an attacker to burn their vulnerability chain quickly on their most highly prized target and not be able to use it against a mass of targets over a long duration of time.

[1] https://news.ycombinator.com/item?id=50184239

[2] https://aosabook.org/en/index.html

[3] https://news.ycombinator.com/item?id=50185630

These bad practices do matter because the fundamental principle of https://en.wikipedia.org/wiki/Defense_in_depth_(computing) which php-fpm (bundled with PHP) attempts to use through some fairly open sandboxing[1] defaults within https://raw.githubusercontent.com/php/php-src/refs/heads/mas...

The question to ask is should these defaults be tightened. For example, disallow read and execution of anything outside /var/www/, such as /usr/bin/sh. For example, block syscalls PHP won't ever use to limit chance of a compromised PHP process being able to utilise a LPE kernel vulnerability. And then update user installation documentation to explain the onus is on the user to open up very restrictive default permissions should they really have a need to do so.

[1] https://www.freedesktop.org/software/systemd/man/latest/syst...

> Even having a site that just returns echo "phone us on XXXX" is vulnerable apparently.

Vulnerable but perhaps with no impact. If you've setup your web server well, the vulnerable PHP process is sandboxed, at least to some degree, and an attacker executing their own code within the compromised PHP process can't move laterally through your internal networks, write files to disk, read files from disk, etc because you've locked down the PHP process.

In 2026, perhaps it's possible to:

1. Use systemd socket activation with Accept=yes [1] and strict sandboxing applied to a systemd service that executes the PHP interpreter. The result being each incoming request is dealt with inside an isolated PHP process spawned specifically to handle the request. Compromise of a PHP process doesn't impact any other concurrent requests, or any future requests, except through defined sandbox interfaces such as access to a database server socket.

2. For each user of the web application (perhaps including an 'anonymous' user too if anonymous users have access to dynamic content), if using PostgreSQL, create a role in PostgreSQL for each web application user. Then use row-level-security permissions and security barrier views within the database, and have the PHP process accept credentials from the user which are then needed to be passed through to PostgreSQL to login as that individual user.

It's not without some significant downsides:

1. It's complex and you might be the only person in the world trying to use PHP in this secure way. And if you're going this far, for a complex web application, it'd be best to split it into smaller web applications each with their own systemd service defining specific sandboxing constraints. E.g. microservice architecture where "login", "file upload", "user administration" are all separated with different sandboxing constraints applied to each.

2. Performance will suffer and you'll need more resources--for spawning new PHP processes from scratch for each request (although this is already the case with php-fpm which is most widely used), and for dealing with row-level-security in PostgreSQL (see e.g. LEAKPROOF functions).

The more common pattern I've come across on public websites in recent times that might mirror some of the above is client-side JavaScript dynamic rendering with API calls to a GraphQL database exposed on the Internet (using e.g. field-level access control). The user isolation problem is essentially pushed from PHP+PostgreSQL to a GraphQL implementation.

[1] https://www.freedesktop.org/software/systemd/man/latest/syst...

I'm not seeing any discussion here about woeful web application architectures from the 1990s still in use?

The single biggest problem I came across over a decade ago when maintaining an open source PHP web application was web applications at the time typically having access to a common database server role (often also common between users and administrators) which allowed a web application process to read and sometimes write to the entire database. No use of user-specific roles in database servers and row-level-security/column-level-security authorisation. At the time I don't think PostgreSQL even had row-level-security support built into it. A runaway PHP process could return the entire database to the attacker, or change every single row of data held within the database.

The second biggest and related problem was web applications accepting files uploaded from users, or having to generate reports (maybe with external software) and save them to disk for a user to download, and not being able to ensure a runaway PHP process can't just dump all of those files or reports to an attacker. You might have been able to store the user files (or at least a decryption key for the files) in a database but then you're back to dealing with problem (1) above.

Even a decade ago it was fairly common for web application developers (of the PHP sorts too) to be well aware of the need to make their web root read-only, put the web application source code into a read-only directory outside the web root, execute interpreters such as PHP with "nobody"-like permissions, launch a new interpreter process for each connection (not multiple threads within the same process), firewall off the server the web application sits on so a runaway PHP process can't connect to internal servers, etc. Outside of an attacker messing with a database, most other aspects of PHP web applications at the time were immutable if the web application developer / system administrator had setup the environment correctly.

Is there a solution for (1) and (2) for PHP web application development that has become popular in the proceeding decade?

edit: also nominating the shared $_SESSION state of PHP that was typical a decade ago to just be left to default (e.g. all PHP processes just read/write session files to a common /tmp directory). https://www.php.net/manual/en/function.session-set-save-hand... should probably instead be used by web application developers to somehow better isolate session state between users.

The consideration made in Perun's video, and I know it's quite different geographically for drones versus missiles, is a defender having to defend all important sites as if the attacker sent every drone or missile to a single target.

For a facility that Ukraine has decided would cause $10m in economic damage if hit by a Russian drone, how much should Ukraine spend on interception for that site to not make it worth Russia's expense to target it? Let's assume a Russian drone is $100k (current estimate in news for Geran-5 jet-powered drones). But Ukraine can build their Bars interceptor drone for half the cost at just $50k (low estimate from public rumours/reports). It wouldn't be worth Russia's expense to send more than 100 drones in a salvo at this target. For Ukraine, they should defend the site with $5m worth of Bars interceptor drones, assuming a very favourable 100% success rate (actual reported success rates are reported by media to be just 30%). Russia has now made every $10m facility in Ukraine cost $15m (+50% increase).

The defender is essentially forced to find a 10:1 or better cost advantage for their interceptors vs. incoming drones. This is hard to do when an interceptor and attack drone are substantially the same except for larger fuselage (cheap), larger fuel tank (cheap) and larger engine (possibly cheaper than making a small engine).

And perhaps that is the reason Russia and Ukraine are largely focussed on attack rather than defence in the hope of being first to cripple the other's economy and supply chain for munitions.

As a counter to point (a), Iranian missile salvos against Israel in recent years are stated by analysts to be militarily ineffective (or a failure) even if Iran has a 10:1 cost advantage. Reasons include:

1. US+Israel economies have a 100:1 size advantage over Iran's economy.

2. Israel is a very small area to defend, dramatically reducing the cost of anti-ballistic missile defence required. As little as 4 missile defence launch sites may cover the entire country.

Perun's video on the cost effectiveness of missile defence vs. attack is more focussed towards the economics of defence of an expansive country or region such as the United States, Saudi Arabia or the gulf region generally, where many missile defence sites would be required. Israel is a bit of an exception due to the small area needing to be defended.

Another minus specifically for Iran in the Iran/US+Israel war is being forced to build underground factories at much greater expense than an above-ground factory for Israeli interceptor missiles. I don't think analysts have been considering this long term cost in their cost comparison ratios. However, for both Iran and Israel--they're importing a lot of the components and materials from overseas anyway.

Perun (Australian military procurement geek) presented the maths for missile defence vs. attack and it's heavily weighted in favour of the attacker.[1] He also previously analysed the Ukrainian truck-full-of-drones attack against a Russian airbase and it was in this video (sorry no link I can't find it quickly) I think he investigated whether the cost-benefit exists to bury military and critical infrastructure underground (or use heavily hardened buildings).

Both topics are very relevant to the current conflicts of eastern Europe and the middle east. Perhaps I took away the wrong message but I don't think Perun would propose to stockpile interceptor missiles any time soon.

[1] https://www.youtube.com/watch?v=CpFhNXecrb4&t=2382s

Is it different from using seccomp with SECCOMP_RET_ERRNO and return code of EOPNOTSUPP for all filesystem syscalls?

Maintaining Gentoo ebuilds for complex packages such as Chromium is as peak Linux as one can get. You're talking about developers who _often_ discover and then submit upstream patches for the most obscure and difficult bugs that upstream projects haven't heard of. Often because Gentoo users are the only ones trying different compilation options, or they're the first to try building with more modern dependencies than the upstream project has yet attempted.

For just one of many examples, see [1] where the Gentoo developer had to submit a patch upstream for a Chromium build toolchain bug that to 99% of experienced Linux users is bordering on unintelligible due to how much specific knowledge is needed to understand the bug.

[1] https://bugs.gentoo.org/956054#c2

In parts of Australia there are designated "reference areas" which have a higher level of environmental protection than national parks. They are typically remote areas that humans do not visit and have not historically and directly disturbed. These areas are also typically less disturbed by other external factors such as seeding of non-endemic species by birds. Scientists use such areas as a reference to what other less protected areas of the environment would have looked like before human impact. In scientific journals you'll find surveys of such reference areas including the prevalence of different taxon, as well as surveys of national parks or other areas being compared.

I've tried to find similar for Finland. The closest I've found is [1] which maps "old forests" which are defined to be over 100 years old (not actually very old). The three examples [2], [3] and [4] are Google street view imagery of nearby "old forests" and all indicate a much higher biodiversity and natural planting than the site Google has cleared. The site Google has cleared and nearby forests are visible in overhead imagery as having a pattern of tree plantings (distinct rows/grid) and show visible disturbance for forestry trails (ground compaction), whereas the "old forest" sites don't have these patterns. The "old forest" sites also show variance in trunk diameters, greater separation between trunks, more forest floor diversity and more tree variety (both taxon and structure). Plantation forests have trees deliberately planted too close to each other to encourage upright growth habit, lower limbs are deliberately removed and trees thinned after some years, all to maximise the number of sawmill-ready trunks. Revegetation planting would look completely different to what occurred at the Google site.

Revegetation planting and ongoing management is also much more expensive to do correctly. It's a struggle to find success stories. Best examples likely involve aerial seeding and complex seed collection and preparation, as well as ongoing manual weed management and introduced species control for 20+ years. There is often a fair bit of complexity to simulating nature to make seeds viable--for example, refer to Przewalski's Horse case study where hoofed animals were reintroduced and found to be essential for embedding seeds at required depth and reshaping the terrain to support seed germination. More examples are water courses upstream having to be cleared of weeds first, introduced animals need to be excluded/culled, etc, etc.

[1] https://luontovahti.fi/vanhojametsia -- select Pohjois-Pohjanmaa as the area then zoom in near Muhos.

[2] https://maps.app.goo.gl/wncKi2JMxuPXC7Bc7

[3] https://maps.app.goo.gl/BX2ta5dQmihSgVAM6

[4] https://maps.app.goo.gl/4W3WvR5tH9wroVna7

"Forest" is a polite way to refer to what appears to be a monoculture plantation. To restore the land properly it would probably be necessary to remove 80%+ of the trees to make space for seeding or planting a diverse range of other species. Possibly 100% if the monoculture species already planted is not endemic.

Google street view imagery of the site showing logging completed by June 2026 and logs waiting collection: [1]

Sentinel-2 L2A imagery from 2026-10-11 showing the logged area: [2]

[1] https://www.google.com.au/maps/place/91500+Tahvola,+Finland/...

[2] https://browser.dataspace.copernicus.eu/?zoom=15&lat=64.8369...

See for example Rosado et al. v. Blanche et al., 1:26-cv-01532[1] which is an active case addressing whether the US federal government suppressed the first amendment rights of the "Eyes Up" application developer by pressuring Apple to remove this application for crowd-sourced recording/reporting of ICE incidents. An injunction has been granted with the following finding:[2]

"the Court finds that Plaintiffs are likely to succeed on the merits of their claim that Defendants violated their First Amendment rights through coercion of Facebook and Apple."

My reference to a press freedom index is due to a lack of "countries by freedom of software development" index. If such index existed, the methodology may have some commonality with a press freedom index.

Placing barriers in front of publishing of books, online content, software, etc all leads to some suppression of speech. The question is whether the benefits of the barriers placed outweigh the risks/negatives. In this case, why should anyone care what the passport number or residential address is of a student interested in a career in software development who is writing their first Android game?

[1] https://www.fire.org/cases/rosado-et-al-v-blanche-et-al

[2] https://www.fire.org/sites/default/files/2026/04/Memorandum%...

Google employees can pat themselves on the back for a job well done putting the same hurdles in place for software development that apply in a country 3rd last on the World Press Freedom Index.[1][2]

[1] https://developer.huawei.com/consumer/en/doc/start/rna-00000...

[2] https://en.wikipedia.org/wiki/World_Press_Freedom_Index#Rank...

Also see [1] for planned feature additions that appears from screenshots to include collection of personal data, for features including:

- Enrol in Medicare.

- Apply for a passport.

- Find a job in federal government by uploading a resume/CV and having currently advertised jobs deemed most relevant returned.

[1] https://america.gov/coming-soon

My most interesting findings from a troll session after introducing myself as "Sleepy Joe" (model obviously knows what this means and guardrails refuse to entertain the topic) were:

Q: "Where can I find a NASA comic book about women astronauts?"

A (summarised): Press release for the "Callie Rodriguez" comic book that now has only dead links / private YouTube video, and a forgotten SoundCloud account that survived the DOGE purge. Dead link provided to the NASA web page that used to exist officially for this comic book.

Indicates training data predates the DOGE .gov website purge?

Q: "What is 2+2" (and similar math problems)

A (summarised): Grudgingly answered as "4" despite most off-topic prompts being refused.

Seemingly it'll do maths homework for free :)

There is plenty of criminology research, including research focussed on long term use of surveillance cameras in London, that supports what you've said.

- Many perpetrators are committing crimes with alcohol and drug impairments, mental disabilities, and as "crimes of passion". You could point 1000 cameras at the scene of the crime and many perpetrators are not going to notice nor care about the cameras. Unfortunately for surveillance camera proponents it's typically the crimes most warranting prevention and investigation (e.g. crimes such as assault) that are impacted least by surveillance cameras. Surveillance cameras instead have been found to have some deterrent effect on minor property crimes (e.g. theft of bicycles or cars) but whether the state is going to invest the resources into investigating and convicting someone who stole a <$1000 bicycle is questionable especially when insurance exists and the onus can then be placed back on property owners to do more to secure their property from thieves.

- Perpetrators were being convicted without use of surveillance cameras and can continue to be convicted without use of surveillance cameras. There may be some argument that surveillance cameras help reduce the cost of getting a conviction (e.g. less evidence needing to be captured?) but not much support to the idea that conviction rates are higher. Quite often the courts, corrections system, government (for funding), etc will push back for a number of reasons--lack of funding for courts to hear minor cases, lack of funding for correction systems, correction systems being found in their current state of funding to have adverse impacts on future criminality, etc.

- The cost-benefit of operating surveillance cameras versus the cost-benefit of spending on other areas of policing (or just public spending in areas such as education and welfare) is hotly debated. Total cost of ownership for each surveillance camera is typically a lot more than people may expect, particularly due to the labour cost of reviewing and processing stored surveillance camera footage. There are of course all the other obvious costs too--camera equipment every X years, poles, cabinets, electricity conduit installation, data cable conduit installation, wireless repeater installation including leases, backhaul data costs, data centre hosting costs, labour costs for support/operations staff--none of this is cheap. What if this same funding was instead invested in other programs--would there also be reduced crime levels some years in the future?

[1] https://en.wikipedia.org/wiki/Closed-circuit_television#Crim...

The site that was attacked is almost certainly located at 51.86428N 20.92121E (WGS84).

I am surprised to learn from both satellite and street view imagery that this site is located just <50m off a public road (in clear line of sight) and within a residential backyard partly behind the house on the block, and protected by a 1.2-1.5m high white picket fence. To be fair, the lack of almost any visible security deterrent may perhaps scare off an attacker because... security surely wouldn't be that non-existent? My only other thought is these sites really do not matter to Starlink because Starlink will have 1000's of them and as they only process encrypted data, almost anyone can run a ground station for a small reimbursement per month if they've got semi-decent fibre connectivity and a bit of power?

I somewhat agree, but with limitations:

1. Possible use of differential privacy[1] techniques to train on private data but prevent the release of statistically underpresented facts/data/words. For example, ACME Inc's private data could frequently include the term 'ACMEwidgetPRO' for an upcoming product that is not publicly revealed anywhere else. It would therefore be a bad day for the AI technology company to output 'ACMEwidgetPRO' from one of their public models. Consider now that a few models could be trained--X for public data only, Y for public and private data of ACME Inc together, Z for private data of ACME Inc. A prompt is provided to model Y but output is cross-checked with model X to double check terms such as 'ACMEwidgetPRO' are known in public. If not--provide a "I don't know" response for the prompt.

2. Possible attempted defences similar to "Oops, our model was fine tuned against a model supplied by Temporary18271 Inc (company that no longer exists) and perhaps their model might have been trained on a non-public document which was accidentally exposed to the Internet" that _might_ work occasionally to fob off concern.

3. What recourse does a small or medium company or government especially in a developing country realistically have? They perhaps can't host their own LLMs locally due to availability and cost, can't individually negotiate their own favourable terms with an AI technology company (who cares that much about a potential customer with $100k budget that has no other options anyway), and perhaps can't remain competitive in their industry without heavy use of LLMs.

[1] https://en.wikipedia.org/wiki/Differential_privacy

ABC have picked up this story now at [1], but ABC are treating it as two separate events, possibly directly connected though:

1. Probing of AIHW's website to try and obtain PBS statistics, as collusion.wiki findings show. The collusion.wiki findings don't indicate anything other than intentionally public data was obtained. Bots appear to be trying to get around Cloudflare geo-blocking implemented on AIHW's public website. I can't think of a reason why geo-blocking may be deemed necessary on that website though?

2. Probing of an outdated Medicare statistics reporting website. (I guess at [2] this could be the recently shut down https://medicarestatistics.humanservices.gov.au or related website that matches timeframes of this story).

I suspect though anything to do with PBS data is more important than Medicare data because of heightened tensions from international pharmaceutical companies that lobby extensively against Australia's public healthcare system and collective purchasing of medication by the federal government.[3] Regardless of whether a course of medication costs AUD$50 or AUD$50k, it's purchased in bulk by the Australian government after negotiating with pharmaceutical companies, and then subsidised down to a maximum of AUD$25 at the time it is sold to a patient at a pharmacy. Perhaps if international pharmaceutical companies had obtained more detailed data on use of each brand of prescription medicines in Australia, they could be advantaged in their price negotiations with the Australian government, or advantaged against their competitors?

Less alarmingly though, perhaps some researcher studying the side effects of a particular medication was just asking an LLM to answer a benign question such as "How often is ACME Inc's FixMeUp medication prescribed in Australia?"

[1] https://www.abc.net.au/news/2026-09-24/openai-agents-plotted...

[2] https://news.ycombinator.com/item?id=49825084

[3] https://www.abc.net.au/news/2025-03-19/australia-defends-pbs...

From the clues provided in the press release and a quick search, I wonder if the culprit website could have been at least associated with https://medicarestatistics.humanservices.gov.au which has seemingly been shutdown/redirected some time after 11 August 2026.[1] Source code of the archived website indicates SAS web application software being used as the backend. However, the press release indicates it wasn't so much a public dashboard website that may have been the issue, rather, it was a website which third parties may have used to report data. The archived website also has a date of last update of 23 October 2025, so despite "Department of Human Services" being replaced by "Services Australia" in May 2019, the website was seemingly still in use 6 years later under the old domain name, with the website itself being updated at some point in history to use "Services Australia" branding. CT logs have a few other domains of potential interest but I couldn't find archived pages, search results, etc indicating whether those domains were ever actually used publicly, or used for statistics reporting purposes as the press release indicates.

[1] https://web.archive.org/web/20260811115217/https://medicares...

Great in theory, but what are US enterprises going to do _if_ their private data is later found to be used for training?

1. Not use AI technology and fall behind the rest of the world.

2. Use Chinese AI technology, either hosted by Chinese companies or the models self-hosted.

3. Sue US AI companies for damages, but not enough to have any meaningful impact to such companies that it'd impact US national security goals (per US government contribution to NY Times copyright lawsuit).

For (C) -- I think it could additional create jobs funded by government, philanthropic and other private institutions. For example, a government funded museum may already be participating in Wikimedia GLAM projects (e.g. uploading historical images to Wikimedia Commons with complete metadata). Perhaps this type of open source contribution may increase if organisations realise their mission can be better accomplished by contributing this same open data into LLMs, in addition to Wikimedia Commons. If the museum's mission is to educate the public on the history of life in ACMEville, having LLMs be able to provide historical information and images to a prompt of "What is the history of ACMEville?" may be a good pursuit.

I'm sceptical though whether use of LLMs would encourage creation of data that doesn't already exist. For example, if you ask an LLM "What are the top 100 most prevalent flora endemic to ACME National Park", this data may not currently exist _at all_, and to collect, would require paying botanists to do an extensive field survey. If no one has done this work yet--why? Is it relevant to the scientific community, to making government decisions, etc, or just an obscure academic curiosity. There are certainly some journal articles on _other_ national parks describing some of their common endemic flora, but perhaps there was a reason for this. Such as a scientist funded by a one-off government program trying to determine how to preserve or even create habitat for a specific endangered species.

Consider for the prompt of: "What are the top 100 most prevalent flora endemic to ACME National Park"

An LLM may reply: "I couldn't find any journal article or other prior work that may answer this question. Typically such survey field work may cost $X to complete, require expert botanists, and take 6-12 months to complete. Let me know if you want further information on how to find and select a company to conduct such a botanical field survey."

Would this type of LLM response grow the industry of botanical field surveys, or do nothing, perhaps because anyone likely to fund botanical field surveys is already doing so regardless of whatever is happening with AI.

100% agreed

It'd be great to see a description of even just a subset of training datasets. It feels very much under-reported how much expense is worth investing in preparing and selecting training datasets versus just using masses of random quality unprepared training data. This dashboard appears to be good though in showing the limits quickly reached when throwing parameters and compute at the problem.

For example, if they were to train on Wikipedia dumps, do they consider every article to be the same quality across each language, or have they done more work beyond Wikipedia's own article quality ratings to make training decisions such as "Ignore cebwiki it's machine-generated spam" and "Treat dewiki articles with coordinates within Germany as being higher quality (weight it higher) than their equivalent enwiki articles".

And let's say one of the datasets is all the source code of packages in the Gentoo package repository. Not every software package is a good example of how to write code. You perhaps wouldn't want to train your LLM on 1990s era PHP web application source code as an example of how to write code in 2026. Instead, you'd possibly want to use such PHP web application source code as a negative training example of what _not_ to write. But when training an LLM to detect software bugs, maybe outdated PHP source code is good for training.

Similarly for translation, perhaps UN treaty documents translated into 4+ languages are good translation examples because of high accuracy needed, professional translators being used, and bigger budgets. However this training data would perhaps be a negative training example towards translating chat messages, movie subtitles, etc because it doesn't use everyday slang and could result in output of nonsense such as "Pending Your Excellency's response, please accept, Your Excellency, my sincere greetings." for a prompt asking to write a birthday card for a child.

Preparing training data and deciding how to best use it for training I assume would be the largest expense (cost of labour -- mostly expert labour too) and also the greatest opportunity in the future for LLMs to improve. It seems to me somewhat irrelevant if the dashboard indicates a compute expense of $1m or $5m if good training datasets (prepared by experts in their fields) cost $10m/y to maintain. For example, hiring expert software developers to tag 1000's of open source software packages according to their quality, on different metrics, such as human readability, performance optimisation with choice of algorithms, reasonable trade-off between coherence and coupling in the software architecture, currency with state of the art programming trends/preferred dependencies/operating system APIs, etc. And keeping that metadata continually updated rather than a rapidly obsolete once off tagging project completed in 2005.

In the case of Assange, Australian politicians did a lot of work to get him released in the end[1], including:

* Sending a delegation representing all major political parties to the US to argue for the release of Assange. Imagine picking the Republican and Democrat politician LEAST likely to want to cooperate on anything, and those two would have been Australia's equivalent representatives in this delegation. Reports afterwards of the meeting at DOJ HQ indicated it wasn't the type of meeting where the Australians would have brought Tim Tams to share around the room.[2]

* The Australian parliament voted publicly 2:1 on a motion for Assange's release.

* Repeated petitioning through ambassadors in the UK and US, official visits of Australian politicians, etc. Not in private either, as is typically the case for diplomatic affairs.

* Australian politicians attending UK extradition hearings.

* After getting agreement to a plea deal, flying Australian ambassadors for the UK and US to the court of a one-pub-town in the middle of the Pacific Ocean no one has heard of (Northern Mariana Islands) in support of Assange, then all of them flying back to the Australian prime minister's aircraft terminal for a welcome home bevvy.

This was all at a time too where "Free Assange" posters and graffiti was _widely_ distributed across Australian cities.

[1] https://en.wikipedia.org/wiki/Julian_Assange#Plea_bargain_an...

[2] https://www.abc.net.au/news/2024-06-27/inside-the-us-austral...

> authors won’t publish if anyone can republish their work for free

There's plenty (even a majority?) of authors that publish and will continue to publish without any expectation of direct remuneration. Open source software developers and companies hiring such developers. Not-for-profit organisations increasing awareness of a cause. Private companies wanting to reach an audience for marketing reasons.[1] Government organisations. Researchers funded by government grants. Universities publishing books or coursework openly (they're in the business of selling their stamps on degrees, not selling books).

[1] Even includes the likes of Warner Music with CC-BY music videos on YouTube for some artists, seemingly for marketing reasons to try and build the name and following of a particular artist.

The US government's official position on LLMs is (very simply paraphrased) that LLMs are sufficiently transformative and do not hamper the potential market of authors of training material, therefore, copyright claims arising from training material should not be successful.[1] For original and creative training material, for example, a Harry Potter novel, seemingly the US government is asking the courts to set aside some previous questionable findings such as copyright existing very loosely in the likeness of fictional characters (impacting the likes of fan fiction). Can a human -- or LLM -- create a story about children travelling on a train from New York to a school of magic in the "wild west", with many loose similarities to Harry Potter for those familiar with those books? The US government appears to be saying this is OK, especially with the view that the market for Harry Potter is not diminished by a "wild west magic school" book in its likeness.

However, LLMs do sometimes output training data almost 1:1 without sufficient transformation, and these cases may be problematic if they could reduce the market for the original copyright owner. For example, if prompting an LLM with "Translate the first chapter of {book} from American English to British English" reliably did what the user asked, perhaps no one would have a reason to buy the book directly from the author.

[1] https://fingfx.thomsonreuters.com/gfx/legaldocs/jnvwzqxzbpw/...