By Adam Brock

No Means Yes: The Rogue AI Agent That “Didn’t Accept No for an Answer”

My take on the Medicare “rogue AI agent” story, the technical details behind it, and how it was framed.

It was a “world first”. A “rogue” AI agent had “broken loose”, “infiltrated” the Australian government, “scaled the fence”, “outsmarted” its defences, refused to take “no” for an answer, and escaped “human control”.

It was a “wake-up call”, a “real-life example of a science fiction scenario”, and a cause for “extreme concern”.

The machine’s sinister objective? Researching public medical spending.

A note on what’s still unknown, as of 27 September 2026: this is my reading of the evidence available so far. I’m still waiting for a fuller public technical account from OpenAI and the findings of the government’s forensic investigation. The exact access path, the files reached and whether the agent exceeded its instructions remain unresolved in the public record. Further evidence could change my interpretation.

Update: 10 November 2026

Well, six weeks later, the ABC is reporting that OpenAI appears not to have broken any existing law in the supposed Medicare “hack”. The Australian Federal Police has not opened a criminal investigation or prosecution because it received no formal report of a crime. Albanese has also confirmed that no personal information was accessed. Compare that with the original story about a “rogue” agent “infiltrating” Medicare and refusing to take “no” for an answer.

And what came of this supposedly unprecedented threat? The government has accepted all 18 recommendations from its rapid review and announced mandatory AI standards, rapid incident reporting, and a new regulator with the power to impose significant penalties. It is even considering a new criminal offence for companies that fail to prevent harmful AI conduct. The same recommendations also tell government agencies to get back to basic cyber security and shut down or segregate systems that don’t need to be connected to the internet.

That is the argument I made in September playing out in front of us. An agent researching public health statistics was presented to the country as an alarming Medicare break-in, and that incident has now become a political case for regulating frontier AI companies. The ABC’s legal reporting does not tell us every technical step the agent took, but it makes the gulf between the original rhetoric and what has actually been established impossible to ignore. It also gives real weight to my hypothesis about why the story was framed that way in the first place.

Why did I write this?

The day this story broke, my parents called to ask if I knew AI had “hacked Australia”. Later, my grandmother called too. She watches the news several times a day and was just as confused and worried. Journalists at one press conference seemed worried too:

It sounds a little bit terrifying, isn’t it? It’s just the start. Should the public be worried?

Most people don’t have the time, interest or technical background to work out what happened. They have to rely on the government’s account. In the examples I cite, official claims were repeated in coverage before key technical details were public, and some headlines grew more dramatic as the day went on.

Watching when and how this event was announced, and how it was reported, is a personal example of the Gell-Mann amnesia effect. I know enough about AI agents and websites to notice when the coverage doesn’t add up, yet I still have to rely on coverage for subjects I don’t know as well.

The phenomenon of a person trusting newspapers for topics which that person is not knowledgeable about, despite recognizing the newspaper as being extremely inaccurate on certain topics which that person is knowledgeable about.

I’m also increasingly worried about the broader mass hysteria around “AI”. I worry that events like this may be used to scare the pants off people with non-stop talk of Armageddon, so tighter centralised control over AI can then be offered as the solution that will save them, all in the name of peace and safety. I can’t establish from the public evidence whether that was anyone’s intention.

For the way this was handled, Kent Brockman from The Simpsons said it best:

Professor, without knowing precisely what the danger is, would you say it’s time for our viewers to crack each other’s heads open and feast on the goo inside?

Yes I would, Kent. No I wouldn’t, Kent.

I hope this is useful as my own opinion on what happened here. This is speculation, not inside information: it’s my reconstruction from public reporting and other publicly available material.

Things didn’t add up from the start

The minute I read the first article about the “rogue agent” from OpenAI that “hacked” and “infiltrated” the Australian government’s Medicare website, I instantly knew that things didn’t add up.

I have been writing code and developing web applications for almost 20 years, and for over 12 months I have used AI coding agents extensively every day to help run my startup, Hypermatic, and many other projects. I also co-run AI Ops Melbourne, a meetup exploring what’s happening at the frontier of agentic AI, so I have a pretty good idea of how websites work, and how these agents work.

On Thursday morning at 7:38am, the morning that the story came out in the news, the first message I sent to my friend group was:

“I will bet that this is not what he’s making it sound like…”

The first article, titled “Anthony Albanese says OpenAI agent hacked Medicare”, said:

This incident occurred in June this year and involved an Open AI agent gaining unauthorised access into the public facing Medicare statistics reporting service portal, which is administered by Services Australia.

No personal information is believed to have been accessed at this stage but investigations are ongoing.

Wait a minute, so the portal was public, and there was no evidence that personal data had been accessed? That didn’t sound like the “Medicare hack” the headlines made me imagine. Of course, a public-facing website can still contain files that aren’t meant to be public, but the distinction seemed important.

The next thing that raised a red flag for me was this comment, reported in the same Guardian live coverage:

Marles went on to say that an AI model was given a “benign task” of researching health and medical statistics and approached four Australian medical websites. Three were only used as a member of the public might, but the agent effectively hacked into the fourth related to Medicare and got the information it was seeking.

Right, so the agent was given a “benign task” of doing research online. It wasn’t “rogue” in the sense of inventing its own mission or acting without a goal; it was pursuing the goal that it had been given. Whether its methods exceeded its instructions is a separate question. It “approached” websites, also known as visiting a website, where it was clearly trying to read the data to pursue its research goal.

What officials said about the “blocks”

Initially, very few details about the “technical” side of what happened were revealed, as the Guardian reported:

We’ve seen an artificial intelligence agent, in an unintended way, access an Australian government website in a way which is unauthorised, and that is obviously deeply unacceptable.

However, as the morning went on, another update from ABC News came out that made me double down on thinking that it sounded way more like bypassing some kind of CAPTCHA than anything else:

But in what is an unprecedented incident, the AI agent broke into Medicare after working around blocks put in place to prevent hacking.

“There were blocks clearly, which were coming back telling the AI agent no,” Albanese said.

“The AI agent found a way around those blocks, didn’t accept ‘no’ for an answer. The model attempted alternate ways to obtain the information that it wanted, and this led to unauthorised access into some other areas.”

The way they kept using the term “blocks” made me wonder whether they were talking about bot protection, although it still wasn’t clear what other restrictions might have been involved.

This was also made clearer when OpenAI gave a brief statement, reported by ABC News, saying:

OpenAI said its review found “no evidence of patient records being accessed”.

“The information accessed included aggregate health statistics and internal file names,” the company said.

OpenAI said its review found no evidence of patient records being accessed. That still leaves questions about what other files the agent reached.

What kind of barrier was it?

The early accounts kept referring to “blocks” without saying what kind they were. I wondered whether the agent had encountered bot protection or a genuine authentication barrier. Later, The Record independently examined the portal’s archived code and found that its statistics service routed visitors to a guest endpoint that required no credentials. That makes the distinction concrete, but it still doesn’t tell us which route the agent actually used. The incident logs have not been released.

Later in the morning, there was a press conference held, which I was watching while also reading pull quotes from the articles being updated about it.

Katy Gallagher also said that $160 million had been allocated in the last budget to improve cybersecurity across Services Australia’s essential infrastructure, and that she had asked whether this work could be accelerated. The portal sat within Services Australia, but the public information doesn’t say what share, if any, of that funding applied to this particular service. What had been done to secure it?

The irony is that calls to tighten AI controls came before the basic questions about the government’s own portal and its security had been answered.

As the press conference progressed, it became clearer that this was a statistics portal, not a system for processing individual Medicare claims:

Finance Minister Katy Gallagher, who also holds the Government Services portfolio, says the Medicare website that was targeted by OpenAI’s agent hosts “publicly available aggregate Medicare and Pharmaceutical Benefits Scheme statistical data”.

They are telling you that the agent viewed, or “targeted”, in their words, a Medicare “website” that hosts publicly available data. That still leaves open what, if anything, it reached beyond those public reports.

“The agent was undertaking a task by OpenAI to conduct internet based research into public medicine spending as part of an internal capability evaluation,” she told reporters in Canberra.

Here, they are admitting that the agent was explicitly given a task by OpenAI to find pages online with public data about medical spending in Australia.

“It’s a public website. It’s most often used by researchers and academics who get that aggregated data about benefit statistics, prescribing statistics to use in their own research.”

Again, by their own admission, they are saying that the agent visited a public website, where the purpose of that website is to provide aggregate data statistics for researchers to use, in this case, an AI agent literally tasked with doing this exact research.

“Scaling the fence”

In later reporting, the description of how the agent got around the site became even more interesting to me. Like using the term “block” earlier, they decided to change the language to “fence” in the press conference:

Marles compared the portal’s security to a fence. “This AI agent scaled the fence, but it did scale it. And the point is, it was unintended. It wasn’t asked to,” he said.

Saying that this agent “scaled the fence”, in my mind, sounded like it may have got around bot protection, rather than cracked some formidable authentication system. But the analogy alone doesn’t establish exactly what happened.

In terms of not being asked to do it, that is what these AI agents do. As Theo quoted from Peter Gostev’s post:

“GPT-5.6-Sol is like a rottweiler who will grab the problem by the throat and not let go until it is done”

I experience this every day using AI agents. They are surprisingly proactive and creative, and that is a huge part of what makes them so useful. They will grab onto a problem and try different approaches without needing me to tell them every next step.

In the consumer tools I use, they also ask for my permission when an action requires approval under my settings. That is different from an internal OpenAI evaluation, where I don’t know what permissions or approval controls were in place.

Persistence is useful, but it doesn’t make every method acceptable. The question is which boundaries this agent crossed, and how.

I use these tools every day, which gives me useful experience and probably some bias too. If the logs show the agent crossed genuine access controls, that deserves scrutiny. So does OpenAI’s handling of the incident.

This is just a basic research task, which is very simple for an AI agent to carry out.

Instead, imagine if you wanted to use AI agents to help you with hard maths, coding, or scientific problems. You would obviously want it to be as proactive as possible in trying new ideas and approaches to achieve the goal you have set for it, in this case, aggregating public information from the internet.

Gallagher said the portal was a legacy website, used mostly by researchers and academics, and was “not in any way related to Medicare in terms of claims, payments, processing individual information”.

She said the site had protections against bots. “Unfortunately, this agent got around that.”

This was another clear “tell” that at least one of the barriers concerned bots and scraping. It still doesn’t tell us exactly which files the agent reached or how it reached them.

Bots already do this

This is not new.

From another article called “99% of My Website Traffic Is Bots”:

On April 22, my site took 3.6 million requests in a single day. They came from 361,844 unique IP addresses, and nearly all of them were in China.

Cloudflare’s Managed Challenge, their invisible CAPTCHA, absorbed 1.18 million of those in the first ten hours. And yet an alarming number of these Chinese bots were passing the challenge.

So the basic idea of automated systems getting around anti bot protections is obviously not new, and is not something uniquely created by AI agents.

What does “writing files” actually mean?

There was a more reasonable article posted online with even more details that gave some more clues into what actually happened here:

The government’s own account creates the sharper unresolved question. Albanese said the model, after being blocked from the Australian data, tried alternatives and “engaged in writing files as well to the internal server.”

This was probably the most serious sounding accusation in the entire story, and at the time I wrote that I would bet a lot that either this never happened in the way it sounded, or that the technical reality was very different from what most people would imagine from hearing that sentence.

Well, we may now have an explanation.

A new technical analysis reconstructed what its author believes was the relevant Medicare portal using archived versions of the website, its publicly available JavaScript, Wayback Machine records, and the underlying SAS reporting architecture. The author is explicit that the exact portal has not been officially identified and that this is a reconstruction, not an account based on the incident logs.

According to the analysis, chart reports generated by the SAS platform automatically saved their rendered images into a temporary folder on the server. Archived examples show files with names like /statistics/temp/gph18MAR25-22-52-54.gif.

In other words, simply asking the website to generate a chart could cause the server itself to write a temporary image file to its own filesystem.

If that is what the government is referring to when it says the agent was “writing files to the internal server”, that description is technically true in the most misleading way imaginable. It conjures up an AI agent breaking into a server and planting files inside it, when it may have simply asked the public statistics website to generate a chart, and the website did exactly what it was programmed to do.

As the analysis puts it:

“None of this requires a software vulnerability. No exploit, no injection, no broken encryption. Only documented endpoints, working as configured, at addresses the site published itself.”

And this wasn’t the only thing the analysis found.

So what did “no” actually mean?

The same technical analysis also gives us a much better idea of what the government’s repeated claim that the AI agent was being told “no” may actually mean.

The portal identified in the analysis had apparently served reports without authentication for more than a decade. After an upgrade to SAS 9.4 in 2025, it gained a login system, but because the statistics were supposed to remain publicly accessible, the site’s built in guest access was enabled.

According to the analysis, requesting the guest endpoint caused the server to generate its own login ticket and automatically authenticate the visitor without requiring a username or password.

As the author puts it:

“The login system wasn’t broken or bypassed. It worked exactly as configured, and it was configured to say yes to everyone.”

At the same time, the human facing website had bot protection in front of it.

So the “no” that the rogue AI agent supposedly refused to accept may not have been an authentication system telling it, “You are not authorised to access this data.”

It may simply have been Cloudflare telling a bot, “You’re not a human.”

On the site the analysis identifies, the reporting system behind that bot protection appears to have been configured to let guests in.

The website’s own code showed how to access the reports

There was another crazy “own-goal” detail in that article.

The site’s own JavaScript apparently published the internal formula used to construct report addresses, including its production path, naming conventions, and references to its test and development environments.

So even the “internal file names” accessed by the agent may be less sinister than that phrase initially sounded.

As the analysis summarises the underlying problem:

“The lock was on the wrong door.”

On the site examined in the analysis, the public facing website had bot protection, while the reporting engine underneath apparently remained accessible to guests.

Importantly, this still doesn’t prove exactly how the agent reached the “non public” material the government says it accessed. The actual request logs haven’t been released, and the government’s forensic investigation is still underway.

But it raises a pretty important question about what “unauthorised access” actually meant in this case.

Did the agent exploit a vulnerability to reach files that were meant to be restricted?

Or was it using the website’s own guest access mechanism, following addressing information published by the website itself, to retrieve information that the government didn’t intend to expose?

Those are two totally different stories.

People were already scraping this data

As of Saturday, September 26, one of the more interesting things I found was a public R script from 2025 for scraping the Medicare statistics website. Its author also posted about it on LinkedIn that year. The README describes how the script retrieves publicly available Medicare item statistics:

This R script scrapes Medicare services data from the Medicare Statistics website https://medicarestatistics.humanservices.gov.au. It retrieves the number of services provided for specified Medicare Benefits Schedule (MBS) or Pharmaceutical Benefits Schedule (PBS) items, broken down by month and state.

It adds a user agent to mimic a browser. It adds a delay between batches to be nice to the server.

This is a totally normal, non AI driven script automating the retrieval of public data from the same statistics website. It shows that people were doing this before the OpenAI incident. It does not show that the script bypassed Cloudflare, or that the OpenAI agent took the same route.

The newer technical analysis makes this even more interesting.

The Medicare statistics portal had apparently been publicly accessible for more than a decade, with report URLs appearing in academic papers, government documents, and other publications.

As the analysis puts it:

“Nobody picked a lock. The gate was open, and it had been open for over ten years.”

Again, this doesn’t prove exactly what requests the OpenAI agent made in June, because those logs still haven’t been released.

But at this point, describing what happened simply as a rogue AI agent “hacking Medicare” seems to leave out basically every important technical detail needed to understand what actually happened.

Facts can wait, let’s regulate!

Whatever the government’s intent, the incident was being used to argue for tighter AI regulation before the technical facts were settled.

Throughout the day(s) ahead, I started to see the rhetoric shift towards regulation, including in Marles’ Today Show interview:

“But let’s be really clear about where Australia stands … Australia absolutely believes that guardrails need to be put in place, that the safety measures around the development of AI need to be well in advance of the capability that is being developed itself, and that extreme care needs to be taken in respect of developing artificial intelligence.”

Greens senator Sarah Hanson-Young went further. The Guardian reported that she wanted the government to “put an immediate stop to the growth of AI in Australia” following the hack. I like the implication that there’s already some great wave of AI growth here that the government can simply stop. What, exactly, would be paused: AI development, deployment, data-centre construction, or all of it?

She also urged the Prime Minister not to waste the moment, as she told ABC’s AM program:

“Big AI needs regulation. And it’s not there yet. And we should put a pause, put a moratorium, before we let them run loose here in Australia.”

My cynical reaction: Never let a (manufactured) crisis go to waste. To be clear, I’m not suggesting the incident itself was manufactured or that anyone coordinated the response. I’m pointing to how quickly it became an argument for stopping AI, before the technical details had been established.

The government’s account describes access it calls unauthorised, but the public information still doesn’t explain what control was crossed or how. Those details matter before this incident is treated as a simple lesson in AI safety.

The ABC also began publishing analysis that connected the incident to the case for tighter regulation, including:

“When rogue AI agent ‘scaled a fence’, it reinforced Albanese’s case for tough guardrails”

One of my friends said they were watching the “7.30 Report”, also broadcast on ABC, where he said they were ranting and raving about the story as if it was our version of Pearl Harbor (his words), saying that we had been “infiltrated”.

In this case, I’m more concerned about the certainty of the coverage than the uncertainty it was supposed to explain. Much of the coverage I saw, here and abroad, not only ran with the “hack” story, but increasingly sensationalised the headlines as the day went on.

As one small example, this was one framing, again from the same ABC analysis, as if the agent was a serial killer that escaped his prison cell and went on a rampage assaulting victims who told him to stop:

“an OpenAI agent broke loose, refusing to be deterred when it couldn’t get its way with the Services Australia Medicare statistics portal.”

With this kind of media hyperventilating about AI day in, day out, is it any wonder that only 36% of Australians were willing to trust AI, according to a University of Melbourne and KPMG study, even before this “hack”?

What the Prime Minister said

The Prime Minister announced the incident on the global stage, guns blazing, while the government’s forensic investigation and OpenAI’s own review were both still underway and important details remained unresolved.

His subsequent United Nations address was reported here:

In his speech, Albanese referenced the revelation on Thursday that in June, an OpenAI bot hacked Medicare and three other federal and state government sites.

Now the story has become four government websites being “hacked”. That conflicts with Marles’ earlier account, which said three sites were used as a member of the public might, with only the fourth described as being hacked. This is the report’s summary, rather than a direct quotation from Albanese, so it doesn’t establish that he personally claimed all four were hacked. But it shows how the story was being amplified in the reporting.

“Recently, an artificial intelligence agent infiltrated an Australian government website. This is unacceptable. And it is why Australia has been leading the way in establishing Australian standards for AI.”

And then, in an Asia Society discussion while abroad in the US, reported by AAP:

Speaking at an event at the Asia Society in New York, Albanese said that the concern was not that Australians’ personal information had been breached.

“This was an AI agent acting independently of its instructions, that is, learning by itself. So it’s a bit of a wake up call, really, about the risks that are there.”

Again, I don’t think either of these accusations accurately describes what happened.

The agent was pursuing its assigned research goal. Whether its methods exceeded its instructions is a separate question.

In terms of “learning by itself”, I don’t know what that is supposed to mean as an extra clarification.

It already knows how to use the web and retrieve information from it (one of the main uses of ChatGPT is just a better version of doing “Google” searches that looks up many websites to find answers for the user), it clearly does not need to “learn by itself” to figure out how to do this every time a user asks the agent to find some information online for them.

So, why?

The obvious non technical question to ask would be, why?

The timing adds to my suspicion. The Prime Minister had known about the incident for several days before announcing it in New York, while both investigations were still underway. He called suggestions of political motives “nonsense” and said the government needed to establish the facts first.

In comments on News24, Assistant Minister for Technology Andrew Charlton rejected claims that the timing was politically opportune. He said the government had done its “due diligence rapidly” but needed to ensure the facts were in order before making a formal announcement. In the same account, he described this as “the first cyberattack by an autonomous AI agent”.

That description needs qualification. On 16 July, Hugging Face disclosed an intrusion it said was driven end to end by an autonomous AI-agent system; OpenAI later confirmed its models were involved. By the time Charlton made these comments, that incident was already public. If he meant the first AI-agent intrusion into an Australian government site, he needed to say so. Whether the Medicare incident should be called a “cyberattack” at all also depends on details the public account has not settled, including what controls were crossed and what the agent actually accessed. This doesn’t prove bad faith, but it is another reason to wait for the evidence before treating the framing as settled.

The irony is that both investigations were still underway when he went public with the dramatic framing. Apparently, enough facts to call it an AI agent going rogue, but not yet enough to explain exactly what it had done.

That leaves me with another question: if the government had several days to establish the facts before its press conference, why couldn’t it and the journalists covering the story put together what I was able to in an afternoon, as other writers I cite here did? One person said he had worked out what happened by putting a few prompts into the open-source model GLM-5.3 at 1pm on the day the story broke:

Why didn’t Australian cyber security researchers or Australian tech journalists work this all out by 10am today and post it? Why did I have to write a few prompts into GLM-5.3 at 1pm?

He later set out his analysis in “The OpenAI ‘Hack’ of Australia’s Medicare That Wasn’t”. That doesn’t prove anyone deliberately misled the public. The government’s explanation may account for the gap, and the timing alone doesn’t prove a political motive. But it is part of why I question the framing.

Beyond my own intuitions, there was a tiny quote near the bottom of an article in the Sydney Morning Herald from internal sources that gives us a big clue:

In Australia, some sources, who did not want to be named because they did not want to cut themselves off from the behind the scenes conversations, said the government was using gentler language in private than Albanese had employed in public.

To me, this supports my suspicion that the incident was being used opportunistically to build support for new AI laws. I think of that as a political psyop, but I have no evidence of a coordinated operation or proof that frightening the public was anyone’s plan. That is my interpretation of the public and private language released so far. A fuller technical account could justify more of the alarm than I currently think it does.

A psyop, pronounced sai awp, is short for a psychological operation, which is a planned tactic used to influence the emotions, motives, thoughts, and behavior of a target audience.

Living here as an Australian, I keep seeing a pattern in the examples I cite: government statements, media coverage and academic commentary can repeat the same framing, then appear to reinforce one another.

Nobody needs to tell this unholy trinity what to say. Shared assumptions and incentives can do the work.

Academics give an argument intellectual authority, journalists present it as consensus, and the government uses that consensus to justify action. What gets presented as independent confirmation can be the same claim travelling between institutions.

Are you going to take their word for it, or ask what evidence sits underneath the agreement?

You can see parts of this process in the coverage.

First, Albanese presented the incident as a reason for AI standards:

This incident is a clear illustration of why we are moving to establish Australian standards for AI

The next day, ABC’s headline presented it as strengthening the case for those standards:

OpenAI breach strengthens Australia’s case for tougher AI safety rules

Then, alongside political calls for consequences, ABC quoted University of Sydney AI evaluation and governance expert Rebecca Johnson:

Any person who broke through to Medicare would be in trouble. A company that does that needs to be held responsible and accountable

That doesn’t prove coordination, or that anyone was wrong. But agreement about what should happen was gathering momentum while the technical account remained incomplete.

The gaslighting will continue until morale improves.

The damage is already done

I think this framing was irresponsible, whether or not the effect was intended.

I worry that dramatic headlines about “rogue” AI agents will stick around even if the technical details tell a less dramatic story. We have seen how hard it is to correct an impression once it spreads.

In Empire of AI, a water-use comparison overstated a proposed data centre’s water use relative to a city by 1,000 times.

The source document had labelled cubic metres as litres, understating the city’s water use by a factor of 1,000. That made the data centre’s maximum permitted water use look roughly 1,000 times greater than the city’s annual residential use, when the corrected comparison was about 1.045 times, or 4.5% more.

As Karen Hao acknowledged in her correction:

As a result the document understated Cerrillos’ residential water use by a factor of 1,000, which made my subsequent comparison between the population’s and data center’s water usage also off by a factor of 1,000.

The author corrected the book, but the original claim had already been shared.

Almost a year since the correction, I still see people cite water use as a reason to oppose AI and data centres, often without that context.

I think this Medicare “hack” story will have the same effect, and the damage is already done.

In the articles and Reddit threads I read, some people were calling for AI regulation and directing their anger at OpenAI. That is my observation of the material I saw, not a measure of public opinion as a whole.

Show us what the agent was told to do, which restrictions it crossed, and what it accessed. If that changes the story, I’ll change my mind. Until then, the confidence of the headlines is doing more work than the evidence.

Whatever the intent, the effect is already visible in the examples I’ve seen: people are frightened, and tighter AI regulation is being presented as the answer.

I’m (not at all) hopeful that, if the final reports are made public and show that parts of this story misled people, the government and media will make just as much of a song and dance about the corrections as they did about scaring the pants off everyone before the full picture was clear.

Written by Adam Brock, with help from Codex to find sources, organise sections, distinguish assumptions from established facts and add contextual links.

Book links are affiliate links. As an Amazon Associate I earn from qualifying purchases.