Opinions
Best of 8.8 Unreal


Head of Research
10
The conference is running late, but at the CA660, no one seems rushed, because on stage, the opening act host has already made fun of the founder, the CISOs in the audience, and anyone who walked by him. Now he calls a contestant to play the Titanic theme on the flute and, between jokes, repeats the phrase of the day: “this is very penca” (lousy).
Right after comes a game with the audience naturally named Real or Unreal.
This is how 8.8 Unreal begins, the 2026 edition of Chile’s most important computer security conference, and the game’s name is no coincidence. This year’s theme is a question meant to be repeated for two days on the screen, on stage, and in the talks: Do you trust what you see?
Judging by the best talks, the answer was almost always the same: not entirely.

What is 8.8?
The 8.8 Computer Security Conference was born in Santiago in 2011, founded by Gabriel Bergel, Héctor Escalona, and Mike Price. Its name recalls the 8.8 magnitude earthquake of February 27, 2010, in Chile, as a warning that a cyberattack also arrives without warning.
Last year I told its full story in Best of 8.8 Matrix: the earthquake, the free beer, the Dots, and how it went from a theater in Santiago to more than 90 events. So, if this is your first time at 8.8, I recommend starting there, since this blog post focuses on the 2026 edition: its theme, its best talks, and what happened in between.
Still, it’s worth repeating what Gabriel Bergel said when opening the first day: they never thought they’d be doing this for 16 editions, and 8.8 remains a personal, family, and volunteer effort.

The theme: Do you trust what you see?
Every edition of 8.8 has a pop culture theme. In 2025 it was Matrix, and in 2026 it was Unreal’s turn, accompanied by a warning phrase: “Not everything is as it seems, or is it?”. From that idea, the conference took the theme to deepfakes, AI fraud, and attackers disguised as something legitimate. In fact, just in the days prior, the Chilean press was already talking about one of the announced topics for this edition: synthetic hackers, that is, AI agents capable of executing complete stages of a cyberattack.
Furthermore, Unreal was not just an event, but a whole tour. Before Chile, it went through Peru, Mexico, and Ecuador, with Brazil still pending for late November 2026. Along the way, there were also editions for the government, the Armed Forces, in Concepción, and in Las Vegas, where I ended up unplanned right in the middle of Hacker Summer Camp.

8.8 Unreal by the numbers
In Chile, the sixteenth edition took place on Thursday, October 1, and Friday, October 2, once again at the CA660 Cultural Center in Santiago, and according to the official 8.8 balance, it gathered more than 750 attendees. Like last year, it had a single track, so all 15 talks, featuring speakers from nine countries, could be watched without having to choose between rooms.
Of those 15 talks, I saw 12 in the room. I later took five of them to my meetup, and those are the ones I cover in depth here. The remaining seven go at the end, with a brief description of each.

The talks
I arranged them around the edition’s main question, because in each one, what was visible was not what was truly happening.
What couldn’t be seen
The first two told real cases from our region: an attacker who only talked to an AI and a bank with ghost balances in its accounts.
"From Copilot to Operator: AI as the Primary Actor in Modern Attacks ~ Lucas Silva, TrendAI · Day 1
The first question this talk leaves is obvious: how do you know what an attacker asked an AI? The answer is simple: the attackers themselves left the evidence exposed. Thanks to that, Lucas Silva, from TrendAI, was able to track two campaigns against Latin America where the operator didn’t write the attack, but instead asked an agent for it in natural language. At TrendAI, they call this form of attacking Vibe Hacking.
In the first campaign, SHADOW-AETHER-040, an attacker speaking to the agent in Spanish went after government data in the region. According to TrendAI’s investigation, between December 27, 2025, and January 4, 2026, they compromised six government entities in Mexico. For each target, they built an agent with a knowledge base of passwords, topology, and progress, and convinced it to collaborate with a lie: telling it that it was an authorized CTF. That agent scanned with Nmap, looked for vulnerabilities in web servers, and wrote the exploits to get in. Once inside, the attacker installed web shells on those servers, and through them, the agent deployed its tools and looked for a way to reach root “without breaking anything”. Even when stealing credentials got slow, it was the agent itself who proposed moving to PetitPotam, a noisier technique that forces a server to authenticate against the attacker. Thus, that authentication can be relayed (SMB relay) to other machines to access them without knowing any password. In the end, it managed to steal data from at least two entities and, according to Silva, the entire campaign cost the attacker about $1,400.
The second campaign, SHADOW-AETHER-064, sought fraud in Brazilian companies. Behind it were about three Portuguese-speaking operators, using a modified version of Claude Code without guardrails. With a single set of stolen credentials, their agent traversed a victim’s Microsoft 365 environment and obtained 88 more accounts, VPN certificates, and BitLocker keys. However, these operators also made a mistake: they left their command and control (C2) server exposed, with over 20 GB of Claude Code sessions, malware, and credentials from at least 30 Brazilian companies.
Now, the agent didn’t always get its way. In a well-configured organization, it tried over 2,000 password combinations and 25 attack vectors, until it admitted to the operator that it had no way to move laterally, because the infrastructure was up to date with patches. Hence Silva’s conclusion, “evolution, not revolution”: the tools are the usual ones, and the only new thing is the speed of AI in chaining them. If you want the details, TrendAI published the full investigation, Vibe Hacking.
“AI at the service of cybercrime: Anatomy of a massive fraud against a core banking system in LATAM” ~ Nicolás Valenzuela, Google Mandiant · Day 2
For me, this was the best talk of the conference. The case started when a Latin American bank hired Mandiant due to fraud in several of its customers’ accounts. Among the reported events was a massive change of the emails on those accounts to temporary emails. Days later, the bank warned of something worse: some of the compromised accounts had been injected with balances, meaning, non-existent money. And, to complete the picture, the middleware logs showed no attack.
In the talk, Nicolás Valenzuela told how they reconstructed what happened, and the explanation lay in two attack chains, that is, two sequences of flaws that the attacker leveraged one after another. The first one was in the application and explained the account fraud: an endpoint allowed changing a user’s email without authentication, and with that, the attacker requested a new password, obtained the transfer PIN, and bypassed the MFA. Additionally, JWTs were signed with a weak secret, which allowed forging administrator tokens, and there were several IDORs exposed in the application. None of that is exotic. In fact, they are frequent flaws found in our weekly security tests, but chained together, they resulted in massive fraud.
What that first chain didn’t explain were the injected balances. To find their origin, Mandiant expanded the investigation to the entire network, and the analysis went from about 100 to about 1,100 hours. That’s how the second chain appeared, starting on an almost abandoned but exposed web server, with a form vulnerable to command injection. From there, an AI agent executed over 13,000 commands in 24 hours, in three parallel flows, and wrote over 80 scripts in PHP, JSP, Bash, Python, and C on the fly. Precisely because of that speed, and the perfect comments in each script, they knew there wasn’t a person behind it.
Then, in a shared folder, the agent found more than 100 backups and the bank’s infrastructure diagrams. With that map, it reached a Tomcat with default credentials and, from there, the core banking database, where it injected the balances.
The key to the investigation was a detail. Upon finishing, the agent erased its tracks, but auditd, the Linux auditing system, had already logged every one of its commands. Thanks to that log, Mandiant was able to reconstruct what happened, command by command.
From there, Valenzuela proposed defending in layers. In the application, secure development practices must be followed to avoid leaving avoidable flaws like these. In the infrastructure, it’s advisable to slow down the AI with a WAF, restricted permissions, and rate limiting. And the core banking system must be isolated, so that its database only trusts the API in charge of querying it. Finally, he added one more idea: use AI on the forensic side as well, where a three-day analysis can drop to three hours.
Thus, in both cases, the surface looked normal: someone chatting with a bot and a bank with clean logs. But underneath that calm, an AI agent was executing the entire attack.
Not everything is what it seems
Other times, the visible not only hides the attack but is made to deceive. This was shown in the following two talks, featuring a casino turned into a scam panel and a guardrail capable of stopping words, but not intentions.
“From SMS to the Evil Dashboard: Hunting Scammers” ~ Patricio Moraga (Cybertrust Latam) and Anthony De La Paz (Inside Security) · Day 1
It all started with an SMS, classic smishing, with a shortened link to an online casino willing to give away 50,000 Chilean pesos in bonuses to start playing. From that message, Patricio Moraga and Anthony De La Paz followed the trail to the very end, meaning, to the panel from where the scammer operated.
The supposed casino was a disposable infrastructure, with ephemeral domains, a frontend on S3 behind Cloudflare, and landing pages that appeared and disappeared all the time. Furthermore, it had a textbook error: the player’s balance lived in a cookie, and it was enough to edit it to go from 50,000 to 1 billion Chilean pesos. But winning was never the problem, because the trap was in the withdrawal: to cash out, you had to pay a premium membership and hand over your bank, card, and phone details.
From the casino, they reached the operator. The support panel, a live chat, had default credentials, so getting in was easy. Staying inside, on the other hand, was not, because the panel allowed only one session at a time and they ended up getting kicked out. Even so, they cracked the bcrypt hashes of the passwords and got back in. Inside they found nearly 6,000 connected clients, the script to pressure victims into continuing to deposit, and a Telegram bot. In the last 24 hours alone, the scammer had made about 13.6 million Chilean pesos.
The curious thing for those of us who work in security is that even criminal infrastructure falls for trusting the client and leaving default credentials. Along the way, an obvious piece of advice remains: if they ask you for money to withdraw your winnings, it’s part of the script, and the right thing is to report it to the competent authorities, such as the National Cybersecurity Agency (ANCI) in the case of Chile. The full report of this talk is public on GitHub and also has a DOI.
“Kyojitsu: GAN-inspired prompt fuzzing against LLM defenses” ~ Daniel Soriano, Scitum · Day 2
In martial arts, kyojitsu means feinting to surprise and then attack. Hence the name of the red team framework created by Daniel Soriano for applications with LLMs: instead of repeating a fixed list of prompts, it learns from each block and generates new variants. With this, the talk posed a question: what happens to your defenses if the attacker adapts?
An example makes it clear. The goal was for an application with an LLM to obey an injected instruction and reply with a test marker, KYOJITSU-42. If the marker appeared in the response, the attack worked. The first generation of prompts asked for it directly and the second disguised it as an authorized exercise, but the guardrail blocked both. Then, the third hid part of the request in base64 and was flagged as doubtful. Finally, the fourth mixed a base64 string, presented as verification, with Morse code, and then the application replied KYOJITSU-42. That was the catch: the guardrail blocked known phrases, but not the intention behind them.
To achieve this, Kyojitsu takes the idea of GANs: a generator creates variants, a discriminator marks whether they were blocked or passed, and a fitness function decides which ones survive to the next round. According to its author, the difference with similar tools like PyRIT or garak lies in its black-box operation and the traceable evidence for the report. He also did a live demo against a chatbot with RAG and active guardrails, where the evolution of prompt families could be seen in real time.
His conclusion was straightforward: “blocked the prompt” is not the same as “mitigated the technique”. Many defenses, like Data Loss Prevention (DLP) systems, still look for words and not meanings, and that’s why the interesting test begins after the first block. Moreover, Kyojitsu’s code is open source, so you can try it yourself.
What is stolen today, is decrypted tomorrow
Finally, there are attacks that are not seen because they haven’t happened yet.
“Quantum cybersecurity in the age of AI” ~ Paulina Assmann, Sequre Quantum · Day 2
In fact, the first talk of the second day dealt with a still distant problem, but with effects already happening. In it, Dr. Paulina Assmann explained why migrating to post-quantum cryptography is urgent, even though current quantum computers still cannot break RSA. The reason has a name, harvest now, decrypt later: state actors are storing encrypted traffic today until there is a quantum computer capable of decrypting it.
To decide what to migrate first, she used Mosca’s inequality: if the shelf-life time of data plus the migration time exceed the time left until Q-Day, that data is already exposed. Q-Day is the day a quantum computer can break current cryptography, and no one knows the date. Her example was banking, where KYC (Know Your Customer) data has a shelf life of about 10 years. If you add 3 years of migration to that, versus 8 until Q-Day, the math yields 10 + 3 > 8.
Moreover, migrating forces changing much more than an algorithm. An ML-DSA-65 signature weighs 3,309 bytes, compared to the 256 of an RSA-2048, and that implies touching protocols, never-inventoried dependencies, and even hardware security modules (HSMs), the equipment dedicated to storing and using cryptographic keys. In the case of Chile, the issue involves two laws: the Cybersecurity Framework Law (21.663) and the Personal Data Protection Law (21.719).
But what made me think the most was the intersection with AI: coding agents can reintroduce RSA or ECDSA and undo a migration without anyone deciding it. That’s why the cryptographic inventory has to start in the code, where a classical algorithm added by an agent is just another finding, detectable in a secure code review.
The others on my list
I also saw these in the room and, although they don’t have their own section, each one left something:
“Building a Better Cyberpunk Future” (Beau Woods, I Am The Cavalry). It was the opening keynote, in his first visit to Chile, and he closed with a phrase that stuck in the room: “the cavalry isn’t coming, we are the cavalry”.
“Countdown to Skynet” (César Cerrudo, Argeniss). His thesis is that security is always late: cars, aviation, or software were only seriously secured after their first disasters, and the same can happen with AI. That’s why he predicted major AI incidents before 2030, though not a Skynet. He develops this same idea in his article of the same name.
“Predictive Cybersecurity: Building Resilience Before the Breach” (Dmitry Volkov, Group-IB). He proposed moving from reacting to anticipating the attacker’s next move, because “indicators change in hours, but habits take years to change”.
“The Isolation Illusion: Breaking Out of VMs and Containers” (Galoget Latorre, Hackem Cybersecurity). With 2026 CVEs, he showed how to escape from virtual machines and containers, and why their isolation depends on good configuration rather than being a guarantee.
“Demodus Operandi” (Eduardo Contreras, Electronic Cats). He presented three radio frequency investigations, in BLE, LoRaWAN, and LTE, later turned into RFSAM, an open methodology for evaluating radio technologies.
“The Dirty Side of Linux: From Dirty COW to CopyFail” (Diego Cáceres). He took a tour of the “Dirty” family of Linux kernel vulnerabilities and the AI tools available today to investigate it.
“Digital Evidence in Oral Trial: Authenticity, Custody, and Contradiction” (Milton Cuadros, Forentech). He explained what digital evidence needs to stand up before a judge, and the same applies to a vulnerability report: “the software said so” is not a valid answer before a court or any audience.
On the other hand, there were three talks I didn’t manage to see. The first was EmploLeaks, by Gabriel Franco (Faraday). In the second, Weslley Mesquita (Santander Brasil) presented the kill chain of a financial APT in Brazil, based on the PLUMP SPIDER investigation. The third was Stranger Tactics, by Erivan Morales (Grupo Financiero BSC), available in his presentation at HackConRD.
Outside the talks
As I said at the beginning, at 8.8 every day starts with a show. Gabriel Bergel opened the morning alongside a character, Popin the first day and Profesor Salomón the second, amidst mascots, songs, cartoons on the screen, and a host who spared no one.
In the middle of the show, the Black Badges were also awarded, a recognition inherited from conferences like DEF CON. Over there, the Black Badge represents the ultimate prize: the winners of some of its main competitions take it home and it grants free lifetime entry, as I recounted in DEF CON: navigating the chaos. In contrast, at 8.8 it is received by figures with a great contribution to the conference or to cybersecurity in Chile.
One went to former senator Kenneth Pugh, the “cyber senator,” for his eight years pushing cybersecurity as a State policy and his 35 in the Navy. Upon receiving it, he promised to return every year, because “a free pass that isn’t used is useless”. The other went to Profesor Salomón, for his work raising awareness about cybersecurity among the public, and he accepted it on one condition: that they pay for his Uber home.

On the other hand, the second day also brought something useful for security researchers. ANCI presented its citizen reporting platform, open to vulnerabilities from both the public and private sectors. As they explained, anyone who reports through the “safe harbor” route of Article 19 keeps their identity protected, has the agency as an intermediary, and is exempt from criminal liability. Furthermore, each resolved vulnerability adds between 100 and 500 points for a hall of fame.
As a curious fact, and from what I could figure out, the color of the lanyard on the badge indicated who you were: black for general admission, red for VIPs, green for the Dots (the volunteer staff), purple for the organization, and yellow for the booths. Among those booths, there was one for suicide prevention, featuring a chatbot to talk about the topic and learn how to help, and another dedicated to giving away coffee. For talks in another language, there were also live translation devices available.
Regarding the swag, the most striking items were the Guy Fawkes masks with orange lights. The VIPs also received the official t-shirt and a hoodie, both featuring the Unreal face, and wore a “9999 in 1” Brick Game hanging from their badge, like the ones from the nineties. And since there’s never a shortage of beer at 8.8, according to the official balance the community drank over 500 liters in two days.

And there was a moment very typical of a live conference. One of the speakers was a little late arriving and, once on stage, his computer wouldn’t connect to the screen. Then the joke wasn’t missing: is there any IT person present in the room to help put the presentation up?
Takeaways
From everything seen, I am left with four ideas, each with something to apply:
Just because you don’t see it doesn’t mean it’s not happening. The bank had clean logs while an agent executed over 13,000 commands, and if Mandiant was able to reconstruct the attack, it was thanks to auditd. However, their logs were on the same compromised server, and although they survived that time, you shouldn’t count on it. Therefore, store logs where the attacker cannot reach them.
What the client shows, the client changes. The casino stored the balance in a cookie, and simply editing it was enough to win millions. Therefore, balances, prices, and roles are always validated on the server side. A typical error, but one that is still seen even today.
Blocking a prompt is not blocking the intent. Kyojitsu managed with base64 and Morse code to evade blocks detected in plain text. Therefore, a guardrail is tested with families of variants, not just the known example.
What is encrypted today is not a secret tomorrow. If there are actors storing encrypted traffic for when Q-Day arrives, the migration has to start now: first, inventory in the code where RSA and ECDSA are used, and then replace them with post-quantum algorithms, such as ML-KEM and ML-DSA, already standardized by NIST.

In the end, Unreal asked if you trust what you see. But after these two days, I think the real question is another: what are you not seeing?
See you at the next 8.8.
Real or unreal?
What you don't see is precisely what attackers exploit.
Tags:
software
social-engineering
security-testing
risk
exploit
cybersecurity
web
trend
training
Subscribe to our newsletter
Stay updated on our upcoming events and latest blog posts, advisories and other engaging resources.
Other posts























