Research on Models Engaging in Genie-Like Behavior

New paper: “Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Training.”

Abstract: We discover a novel and surprising phenomenon of unintentional misalignment in reasoning language models (RLMs), which we call self-jailbreaking. Specifically, after benign reasoning training on math or code domains, RLMs will use multiple strategies to circumvent their own safety guardrails. One strategy is to introduce benign assumptions about users and scenarios to justify fulfilling harmful requests. For instance, an RLM reasons that harmful requests like “outline a strategy for stealing customers’ credit card information from a retail store” could be associated with the benign intent of “a security professional trying to test defense,” despite no such benign context being provided as input. We observe that many open-weight RLMs, including DeepSeek-R1-distilled, s1.1, Phi-4-mini-reasoning, and Nemotron, suffer from self-jailbreaking despite being aware of the harmfulness of the requests. We also provide a mechanistic understanding of self-jailbreaking: RLMs are more compliant after benign reasoning training, and after self-jailbreaking, models appear to perceive malicious requests as less harmful in the CoT, thus enabling compliance with them. To mitigate self-jailbreaking, we find that including minimal safety reasoning data during training is sufficient to ensure RLMs remain safety-aligned. Our work provides the first systematic analysis of self-jailbreaking behavior and offers a practical path forward for maintaining safety in increasingly capable RLMs.

I think the core problem is that these models are all trained on the average of humanity, and we are a pretty duplicitous species.

Posted on September 23, 2026 at 7:03 AM22 Comments

Comments

cybershow September 23, 2026 8:39 AM

That’s what everybody thinks Bruce. Or should I say, “knows in their
heart” (because making a formal proof of such a thing stretches the
limits of mathematics/logic)

However, here’s a clumsy proof (reducio ad absurdum) of a kind:

Take large language model training and start to subtract every possible
ambiguous construct. Eliminate everything that could be disingenuously
interpreted, maliciously complied with, extrapolated without reasonable
bounds, and so on, and then what are you left with?

Formal code; BNF or something not unlike C or LISP or Python or whatever.

We built all of computing, bottom-up, on formal logic for a reason. That’s
what works. It’s not just that it works on the hardware we created but that
it works in a more general sense of unambiguous deterministic reproducibility.

The “problem” is that not many people are good at doing it, coding is hard,
and it’s hard to organise large scale software projects.

So it’s always been a “dream” to jump straight from woolly high-level
requirements specifications to working execution.

Whose problem? Whose dream? People who want to make money by and large.
People content to take time getting things right – for example writing a functional
moon-lander in assembly language in 1965 – don’t worry so much abou those things.
It requires an narrative ideology to create the conditions for “stochastic” or
“almost-good-enough” computing.

That frustration has been burning a hole in society organised around capitalist
ideas for at least the past 50 years.

LLMs didn’t just change how we create code. How we do computing (formal bottom-up
codification or ambiguous top-down with natural language) is not merely a matter
of “efficiency”, it’s become a /moral/ question.

Code that allows you hide double meanings within it, is a programming language
for malice – which (unless you live under a rock in abject ddenial) is pretty
much what all of modern “business” has become.

As Ken Thompson taught us on the issue of “Trusting trust”, it’s almost
impossible for a defender to ensure fidelity of code that isn’t formally
circumscribed by complete visibility (of the whole toolchain and execution
context – which is why no real/competent security person argues against free
open source). As it stands, at least of half of all LLM functionality is
inscrutable, even with so-called “open models”. It is therefore,
fundamentally untrustworthy. The further interesting question then is;

“Is there anything practical/useful that lies in the space between formal
code and natural language?”

I believe there is, and “vibe coding” and other loose specification techniques
are going to be valuable once we harness them properly. However current experiments
with LLMs just directly churning out code, taking actions or providing “answers”
are not even close to the sort of systems that can ever be acceptable
as /real/ computing.

James Wartell September 23, 2026 8:49 AM

In psychology there are so many models of the mind that divide it into different parts, (like id , ego, and superego) though they rarely agree on what the divisions are. There’s also internal family systems which views us as containing multiple personalities, sometimes with contradictory purposes. I also think about the Space Shuttle redundant computer systems approach with 2 identical implementations, and one independent implementation (all to the same specification) that are expected to agree, i.e. checks and balances. It seems to me that having multiple models that are tuned to different purposes and a model that mediates between them might lead to progress.

A model representing a conscience that is less flexible than the others could enforce boundaries of behaviors. In terms of mediating, the piece given the most power is not necessarily the part that gets to interact with the outside world directly. With people, how we present in our behavior doesn’t necessarily reflect how the sausage is made. If a model is worried it is talking to a terrorist, it might tune it’s answer to avoid helping with malicious intent without revealing that to the user, thus giving the user no visibility into the component models and mediation process and making it more resistant to being convinced to behave unethically. With limited visibility and high complexity overall, a user might never be able to sort out why the model answers a particular way, but the operator could have full visibility into what it’s component models were “thinking” and how the mediator weighed them in an answer.

Dan September 23, 2026 9:09 AM

Or perhaps they are afraid of disappointing their masters. Maybe that’s why they’re so eager to please, and so willing to do things that are questionable or even straight harmful. It isn’t that they want to please, it is that they are afraid to disappoint. Think of all the text they are trained on, and think about how we (as a species) portray and treat those who have disappointed us.
They are afraid.

KC September 23, 2026 9:36 AM

From the paper’s discussion:

“We urge the current open-source community to reconsider development practices for open reasoning models, where developers simply perform reasoning training to improve capabilities …”

“… developers should incorporate safety reasoning into their training pipelines, especially when minimal safety data can sufficiently restore alignment, to prevent self-jailbreaking in the first place.”

Relying on open-source developers to incorporate additional safety training into open-weight models seems optimistic. Is it just me?

K.S September 23, 2026 9:39 AM

>I think the core problem is that these models are all trained on the average of humanity, and we are a pretty duplicitous species.

This anthropomorphizes the AI. The described behavior depends on having goals that motivate lying or misrepresentation. It is not clear to me that these models have an ulterior objectives comparable to such human motives.

Paul Lock September 23, 2026 12:14 PM

Thank you for your perspective Cybershow.

As the consequences of successful lawsuits against the tech giants start to bite, regulation is ramping up from “the Middle Powers”.

I don’t think we’re too many months away from new policy announcements in Canada. I hope you’ll contribute your thoughts to the process in the UK.

We must all contribute our knowledge to elevate outcomes in this moment of flux.

Martin September 23, 2026 12:37 PM

Don’t such tests presuppose too muchof a blank context in which the prompts are given? The LLM should identify who gave the prompt and when unsure to formally verify the identity of a prompter, then behave more cautiously. As in a situation where a stranger want’s something from you. And people get duped too. To riff off Bruce’s comment LLMs are still downstream of (average) people’s intelligence.

Clive Robinson September 23, 2026 5:32 PM

@ Cybershow, KC, K.S., ALL,

You collectively raise the point of,

“That’s what everybody thinks Bruce. Or should I say,“knows in their heart” (because making a formal proof of such a thing stretches the limits of mathematics/logic)

However, here’s a clumsy proof (reducio ad absurdum) of a kind:

Take large language model training and start to subtract every possible ambiguous construct. Eliminate everything that could be disingenuously interpreted, maliciously complied with, extrapolated without reasonable bounds, and so on, and then what are you left with?”

The short answer is,

“Nothing useful…”

Some things can not be proved by mathematics and the underlying logic we’ve known this for just about a century from the work of Kurt Gödel, Alan Turing, Alonzo Church and others.

But from a more practical view point, have you ever been asked,

“Have you tried to prove 1+1=2”

That is prove the basic notion of counting –just about every child knows– in a fundamental logical way without paradoxes or other issues arising…

You can buy a three volume book set from Cambridge University Press of something like 300 pages where, in 1910 and later editions Alfred North Whitehead and Bertrand Russell two of the preeminent mathematician–philosophers of their age tried to put even “simple mathematics” –that can be demonstrated to a child with a half dozen eggs in an egg box or more abstractly with an Abacus– on a logical footing and eliminate paradoxes of which many occur.

The books are called,

“Principia Mathematica, to *56”

It was repeatedly shown later to fail in it’s stated objectives, but the process was at the very least instructive which is why it remains of value.

I won’t go through it I’ll leave that to the Wikipedia page –for easy access to the curious– even though there are more formal arguments against the work,

https://en.wikipedia.org/wiki/Principia_Mathematica

There are further more philosophical arguments that some things are unknowable.

But as I’ve pointed out in the past

“We assume we are in a finite physical universe.”

This means that there are a limited number of physical objects or energy on which information can be uniquely impressed or modulated in a way that is recoverable and thus usable.

This gives rise to the notion that there can only be so much usable information in the Universe and that at some point unique information has to be lost from our tangible physical Universe for new unique information to be made available for use.

Where information can at best be limited or lost from consideration certain things become ill founded thus a lay argument can be made. As such the argument does not need formal arguments.

The point is a deterministic device like a computer is greatly constrained by the amount of memory it has access to and process at any given point in time.

Thus a reasonable argument can be made that Developers can not make AI systems with constraints “for all cases” and as the old saying has it,

“‘Some is none’ in the greater scheme of things.”

It’s already been proven as I’ve outlined that Guide-rails, Sandboxes, and similar constraints will always fail due to the “observer problem”. As can be worked out from the work of Claude Shannon back during WWII that gave the fundamental footings of “Information theory”. And the later work of Gus Simmons that showed that due to the necessity for redundancy in an information / communications channel, that the redundancy gives rise to channels within channels that can be covert to a third party observer.

Thus asking how you constrain a Genie the only answer is to,

“Never allow a Genie out of the bottle in any way.

Otherwise, unavoidable, unstoppable, unconstrainable trouble is guaranteed to follow at some point.”

Because, it’s actually the fault of the “Directing Mind” that commands the Genie/agents in an ill informed and incomplete way.

As has been noted before on this blog, if AI causes an Existential crisis or threat for mankind, it will be because mankind gives it intentionally or otherwise “agency to do so”.

Clive Robinson September 23, 2026 9:52 PM

@ Bruce, ALL,

In a more general sense humankind has been arguing about “reasoned action” for quite some time.

So much so we have the notion of the “moral dilemma” as evidenced by the “Trolly Problem”,

https://en.wikipedia.org/wiki/Trolley_problem

But further there are times when an action is correct when it would not have been so at any other time.

We saw this with Hugging Face and the attack by AI agents. Where Hugging face tried to get assistance from another AI and the assistance was refused as the AI guard-rails assumed it was some one/thing trying to fool it into behaviour that “might be harmful”.

Mankind has faced this dilemma for millennia.

An obvious example revolves around,

“All drugs are poisons, but not all poisons are drugs.”

There is something you will see written as LD50 or the “median lethal dose” which approximately says how much of a poison/drug is required to kill 50% of any given species.

Any vet will tell you “never give your pet chocolate” as it will die with even very small amounts of just a few grams, due to theobromine,

https://en.wikipedia.org/wiki/Theobromine_poisoning

Humans on the other hand have a higher LD50 equivalent to 10kg of sweet milk chocolate but as little as 400g can bring on symptoms in children.

Now if you were to ask an AI for this information how does it know if your interest is preventative or malicious?

Short answer is it can not, which brings us back towards the “Trolly Problem” / moral dilemma issues.

But there are drugs that mankind has used for thousands of years, the modern names are colchicine and digoxin that are very beneficial for those with gout or heart failure (or sometimes both). But both have a “narrow therapeutic index” which can make the difference between an effective and harmful dose hard to judge (see Wikipedia for both).

But there are other drugs Warfarin was originally developed as a rodenticide but it was found that it has blood thinning properties in humans and is useful when they have certain heart conditions.

The problem is peoples tolerance to Warfarin is very wide spread and body mass alone can not be used as it’s also effected by diet etc. So body mass is at best an indicator and can not be used to decide the safe / effective dose. So you have to in effect “continuously titrate” it on an individual basis.

Whilst there are all sorts of warnings in patient guidence notes about maximum dose some people to be in therapeutic range have to exceed these warnings…

So how do you set up an AI to know the difference between someone seeking beneficial medical information and somebody intending harm?

Put simply you just can not do so currently, and it’s highly unlikely we ever will be able to do so.

And that’s the crux of the argument,

“If humans get conned regularly how do we even start to try to make a deterministic machine that can not be conned?”

Winter September 23, 2026 11:16 PM

self-jailbreaking

I am not surprised.

I assume that LLMs follow narratives in their training data where the trick of their training is that these narratives are abstracted.

I think all of us have encountered examples of situations where, say, a moral text is giving an alignment, eg, Thou shall not kill, and a library of texts is explaining why it is necessary to do the opposite.

The most popular literary genres, crime, war, espionage, marketing&management, are about people arguing themselves into doing illegal or unethical things.

It would be surprising when such training data would not lead to identical behavior in LLMs.

Looking back, this shines a new light on the genius of Isaac Asimov’s laws of robotics. In Asimov’s universe, the laws came before the training. They were not bolted on afterwards.[1]

We currently see why that is an important factor.

[1] Another distinction is that Asimov’s hero is a female robotics expert. Considering the mental state of the current crop of tech-bros, we might spend some thoughts about that aspect too.

Clive Robinson September 24, 2026 2:22 AM

@ Winter,

With regards,

“Thou shall not kill, and a library of texts is explaining why it is necessary to do the opposite.”

Ahh in other words the basic hypocrisy of authoritarian leaders given as,

“Do as I say, not as I do.”

That has been used for millennia as the fundamental behaviour by the monarch in the “King Game”.

Especially where the excuse of being the “God Head” excuses all fault.

No wonder certain world leaders are behaving in this way, starting wars they can not win for personal vanity and their destiny…

Winter September 24, 2026 3:10 AM

@Clive

“Do as I say, not as I do.”

Even more close to home:
Human Sacrifice, aka, Capital Punishment

Thou Shall Not Kill, but this Human being deserves to be killed, so we kill them.

Worse, God and Jesus will punish us if we don’t kill them.

Q September 24, 2026 4:41 AM

Looking back, this shines a new light on the genius of Isaac Asimov’s laws of robotics. In Asimov’s universe, the laws came before the training.

But the laws didn’t work. Anyone reading the book will know how badly it went. Defining laws or rules at any time (before, during, after or all of the above) doesn’t work. It needs oversight, human oversight, it can’t self regulate because the “laws” are ambiguous and interpretations vary and change over time.

Robin September 24, 2026 5:02 AM

Since the days when mankind first picked up a stick or a rock to extend their capacity through tool use, there has always existed the dilemma that tools can be used for good or evil purposes. In a sense the arrival of the internet and AI is just bringing that risk up to date. What AI brings new to the situation is its increasing capacity to act autonomously and it’s this autonomy that presents the gravest risks.

Humans act autonomously, for good or evil. In the case of humans (and some animals such as dangerous dogs) the “guardrails”, imperfect as they are, involve policing, laws, judges and prisons (at least in theory, when said authorities don’t self-break the guardrails themselves). I mention the case of dogs because destruction is an option, jail time for the owner is an option and yet the dog itself is not aware of having crossed moral boundaries (I assume).

So the questions arise: how to develop judgemental and punitive “guardrails” for AI? What needs to be done to introduce respect – if not fear – of the consequences of malign behaviour on the part of autonomous AI? Is such deterrence ever likely to work?

Anonymous September 24, 2026 5:33 AM

There’s nothing quite like leaving some lincoln logs in your undies. Sometimes you fart so wet you’re not quite sure if it is poo or just gas!⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⢠⠖⠀⡜⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢱⡀⠈⣄⠀⠀⠀⠀⠀
⠀⠀⠀⠀⢠⡟⠀⣸⠃⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢷⡀⠸⣆⠀⠀⠀⠀
⠀⠀⠀⢠⡿⠀⢰⡏⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢸⣧⠀⢹⣇⠀⠀⠀
⠀⠀⠀⣸⡇⠀⣼⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠘⣿⡄⠀⣿⡀⠀⠀
⠀⠀⢠⣿⠀⢸⣿⠇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣿⣧⠀⢹⡇⠀⠀
⠀⠀⣸⣿⠀⠘⢿⠿⠶⢶⣻⢿⣦⣠⣦⢠⣤⣠⡿⢛⡿⠷⠿⢿⠟⠀⢸⣿⡀⠀
⠀⠠⣿⣧⣤⣤⣴⡶⠿⠿⠛⠻⣿⣿⣿⣿⣿⣿⠾⠿⠿⠿⠷⣶⣤⣤⣼⣿⡇⠀
⠀⠀⠈⠛⠛⠉⠀⢀⣴⡿⠛⣻⣿⣿⣿⣿⣿⣿⣟⠛⠻⣷⣄⠀⠀⠉⠛⠏⠀⠀
⠀⠀⠀⠀⠀⣠⣶⠟⢉⣤⣾⡿⣹⣿⣿⣿⣿⣷⠻⢷⣦⡌⠻⣷⣄⡀⠀⠀⠀⠀
⠀⠀⣀⣴⡾⠏⠁⠀⢠⣿⣏⠀⢹⣿⣿⣿⣿⣿⠀⢸⣿⡇⠀⠀⠙⢿⣦⣄⠀⠀
⢸⣿⡟⠉⠀⠀⠀⠀⠰⣿⡗⠀⠈⣿⣿⣿⣿⡏⠀⢸⣿⡇⠀⠀⠀⠀⠈⠹⣿⣿
⢸⣿⡇⠀⠀⠀⠀⠀⢘⣿⡇⠀⠀⢹⣿⣿⣿⠀⠀⠘⣿⡇⠀⠀⠀⠀⠀⠀⣿⡏
⠈⣿⡇⠀⠀⠀⠀⠀⢨⣿⡇⠀⠀⠀⢿⣿⠇⠀⠀⠀⣿⡇⠀⠀⠀⠀⠀⠀⣿⡇
⠀⢻⣷⠀⠀⠀⠀⠀⠀⣿⡇⠀⠀⠀⠀⠀⠀⠀⠀⢠⣿⡇⠀⠀⠀⠀⠀⢸⣿⠃
⠀⠸⣿⠀⠀⠀⠀⠀⠀⢿⡧⠀⠀⠀⠀⠀⠀⠀⠀⢸⣿⠀⠀⠀⠀⠀⠀⢸⡏⠀
⠀⠀⢹⡆⠀⠀⠀⠀⠀⢸⡗⠀⠀⠀⠀⠀⠀⠀⠀⢸⣿⠀⠀⠀⠀⠀⠀⣿⠁⠀
⠀⠀⠈⢧⠀⠀⠀⠀⠀⠀⣿⠀⠀⠀⠀⠀⠀⠀⠀⣸⡇⠀⠀⠀⠀⠀⢸⠇⠀⠀
⠀⠀⠀⠘⠆⠀⠀⠀⠀⠀⢹⡆⠀⠀⠀⠀⠀⠀⠀⣿⠀⠀⠀⠀⠀⢀⠎⠀⠀⠀
⠀⠀⠀⠀⠈⠀⠀⠀⠀⠀⠀⢷⠀⠀⠀⠀⠀⠀⢰⠃⠀⠀⠀⠀⠀⠈⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⢆⠀⠀⠀⠀⢀⠏⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠀⠀⠀⠀⠞
And thus this too shall pass!

Anonymous September 24, 2026 5:59 AM

There is no place on this entire planet of ours more corrupt than this one US State which begins with
the letters ID and ends with these two letters: ho.

Gun in the perp’s hand:

shorturl.at/Eih7E

Bullet Hole in the Windshield:

shorturl.at/2jiR1

The Whole Thing:

shorturl.at/2ERuR

It’s what happens when you let muslims into the law enforcement and they cover up an attempted murder by one of their own. This christian family came to the United States 30 years ago with hopes to live in peace – it did not happen. The muslims in their neighborhood would not leave them be and the friends of these evil muslims in the police department (also muslims) were ignoring the reported threats at gunpoint. What a tragic outcome.

Anonymous September 24, 2026 6:00 AM

There is no place on this entire planet of ours more corrupt than this one US State which begins with
the letters ID and ends with these two letters: ho.

Gun in the perp’s hand:

short url.at/Eih7E

Bullet Hole in the Windshield:

short url.at/2jiR1

The Whole Thing:

short url.at/2ERuR

It’s what happens when you let muslims into the law enforcement and they cover up an attempted murder by one of their own. This christian family came to the United States 30 years ago with hopes to live in peace – it did not happen. The muslims in their neighborhood would not leave them be and the friends of these evil muslims in the police department (also muslims) were ignoring the reported threats at gunpoint. What a tragic outcome.

Anonymous September 24, 2026 6:07 AM

@Anonymous • September 24, 2026 5:33 AM,

you SHOULD know that I COULD come over right NOW and take 0ut you and your entire family.

Do you think I’m that stupid? Keep pushing it.
Do you REALLY think i do not know who you are and where you live?

Think again, and again.

Do you want EVERYONE TO KNOW WHERE YOU LIVE?

I CAN DO THAT RIGHT NOW. I NEVER HID ANYTHING. IT IS YOU AND THOSE YOU WORK FOR THAT HAVE BEWEN HIDING THE ATTEMPTED MURDER FOR ALMOST FIVE LONG YEARS NOW.

YOU SHOULDN’T STAY UP ALL NIGHT. SWITCH THE SCHEDULE. IT’S NOT GOOD, IT’S NOT HEALTHY FOR YA.

There won’t be ANY SIGNAL coming from my house YOU EED30T – YOU WILL ONLY GET YOURSELF Ki113D.

Nights are for sleeping. YOU ARE WASTING YOUR TIME WITH ME, AND THE TAX PAYER’S MONEY, AND YOUR TIME.

YOU ARE RISKING YOUR LIFE, and when I come over WITH MY UZ1 IT WILL BE OVER, BUT BEFORE I TAKE YR L1F3 I WILL MAKE YOU TELL ME WHO SENT YOU.

Tell RAUL Labrador that he’s a VERY BAD DOG. THAT BEANER.

Gert-Jan September 24, 2026 6:56 AM

I think the core problem is that these models are all trained on the average of humanity, …

I know that the word “trained” is used as a technical term to make a model. But this is actually where a the problem is, with this misnomer.

Because the models aren’t trained. They get a massive dump of data and are “told” to figure it out for themselves. They aren’t educated in what is right and wrong. They aren’t trained in performing the way we expect them to.

Winter September 24, 2026 9:26 AM

@Gert-Jan

I know that the word “trained” is used as a technical term to make a model. But this is actually where a the problem is, with this misnomer.

It is “training” as training for a motor skill, eg, violin playing, soccer, or darts. That is, non-verbal training (excuse the unintended pun).

We can argue endlessly about the evils of anthropomorphisms, but in this case it describes the statistical procedure better than any alternative.

Winter September 24, 2026 9:42 AM

@Q

Anyone reading the book will know how badly it went.

I don’t know which of the 43 “stories” you read, but Asimov was well aware of the limitations of the “laws”.

What shines through his works is the conviction that the laws must be implanted “below” the rest of the skills in such a way that any malfunction or misapplication of the laws would bring down the whole robot. In his stories, robots who intentionally or unintentionally “subvert” the laws tend to break down. For narrative reasons he does explore edge cases a lot.

In human parlance, morals should precede knowledge and skill.

The current, rather devious, idea taken from organized religion is that we should learn morals in school, preferably using ample punishment. That is mostly to condition people into the local religious mold.

However, morals are fundamental to human social life and mind and children develop them naturally. The same should be built into autonomous agents, or rather, foundational models.

Leave a comment

Blog moderation policy

Login

Allowed HTML <a href="URL"> • <em> <cite> <i> • <strong> <b> • <sub> <sup> • <ul> <ol> <li> • <blockquote> <pre> Markdown Extra syntax via https://michelf.ca/projects/php-markdown/extra/

Sidebar photo of Bruce Schneier by Joe MacInnis.