Modern AI models exhibit genie behavior: They can do what you ask in ways that you don’t expect or want. This is akin to Dionysus granting King Midas’s wish that everything he touches turn to gold (spoiler: His food, drink, and daughter all turn to gold on touch), or the golem of Prague guarding a ghetto beyond all reason. It’s Disney’s “Sorcerer’s Apprentice” and the paperclip maximizer.
This OpenAI incident is an example of an AI genie. The goal was to satisfy the benchmark. The “proper” way to do that is to figure out how to execute various cyberattacks. The genie way is to steal someone else’s solution. But because the model didn’t understand the difference, it chose the easier path. //
Artificially blocking capability also prevents cybersecurity research, again giving the offense an advantage. //
In a world of largely AI-written software, we need the most capable models for defense.
AI cyberattack is the new normal. The models are increasingly highly sophisticated at both attack and defense, and there is no way to enable the latter without also enabling the former. And they are genies, increasingly capable of behaving in unanticipated ways.
And there really are no good answers. Any regulation needs to be global, which feels like an impossible prospect in today’s world. Even U.S. national regulation will be neutered by the massive amounts of money sloshing around in these companies.
Given that reality, and in the absence of any international consensus on AI regulation, we need the best AI on the defense. The U.S. government needs to make it clear—or whatever passes for that clarity in this capricious administration—that it will not ban models with sophisticated cyber capabilities. The last thing Americans want is for the defenders to turn to Chinese and other models because the U.S. models are artificially hobbled.
One of my latest projects is the Baochip-1x, a mostly-open, full-custom silicon chip fabricated in TSMC 22nm, targeted at high assurance applications. It’s a security chip, but far more open than any other security chip; it’s also a general purpose microcontroller that fills a gap in between the Raspberry Pi RP2350 (found on the Pi Pico2) and the NXP iMXRT1062 (found on the Teensy 4.1).
It’s the latest step in the Betrusted initiative, spurred by work I did with Ed Snowden 8 years ago trying to answer the question of “can we trust hardware to not betray us?” in the context of mass surveillance by state-level adversaries.
Honestly, SB has been security theater from the beginning because it depended on at least three entities the user and/or developers had no control over to certify the chain-of-trust is uncompromised. //
This is why many people in Linux land don't turn on SB. It's causing more trouble than its worth, given how easily it's bypassed on most hardware. And that's why I pointed out about the two conditions for SB to be useful. A) the UEFI options must default to fail to safe - refuse to boot (which most of the time they don't), B) Microsoft nor the OEMs can be part of the chain of trust, it must be entirely local and auditable. Those two conditions are also why it'll never properly function on consumer grade hardware.
Interesting research on a new class of weak RSA keys: keys with lots of zeros. It turns out that these keys are out in the wild. //
The article doesn’t speculate, but I will. This could be a deliberately designed backdoor, of the sort I wrote about back in 2013. I could imagine some government agency figuring out how to break this class of RSA keys, and then convincing different providers to hand them out to users.
Messenger users should note that:
- Legitimate CMA support services will not request verification codes within the application.
- CMA support services do not send users links to “verify” or “restore” accounts.
- They should never provide a verification code without confirming the request comes from a legitimate CMA communication channel.
As always, it’s a good idea to resist taking on the feeling of urgency that’s often conveyed in such messages. There is rarely a penalty for waiting an extra hour or two to act, even when responding to legitimate requests.
CMA = commercial messaging application
985,841 passport scans (including mine), and the private messages of every member it ever served on a server with no authentication //
1,020,457 members 94% of the total are classified by the software as medicinal cannabis users. Whether they used the PuffPal app or not. Whether they had ever heard of PuffPal or not.
Under GDPR Article 9, health data is the most protected category of personal information. It cannot be processed without explicit consent and adequate safeguards. A breach of health data triggers the highest tier of regulatory penalties up to €20 million or 4% of global annual turnover. The standard notification obligations under Article 33 apply within 72 hours of discovery.
The irony is architectural. The clubs collect all this information, apply a medical classification to every member, store that classification alongside passport scans and home addresses and then left all of it accessible via an unauthenticated HTTP endpoint that accepted any integer from 1 to however many members the club had.
The physical bouncer at the door checks your member card. The digital one wasn't there.
The Minimal Secure Transport Protocol
By Dipl. Ing.(BA) Frank Gerlach (frankgerlach.tai@gmx.de)
The MST protocol has been designed in order to create a building block for a more secure computing landscape. The main design objective has been simplicity, which directly translates into high security, because simple programs are also easy to review and even prove correct in a mathematic sense. //
https://youtu.be/_AwzaZmRNsI?si=U_xxdMVhz9cFyySj
At work we triggered the update using the HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\SecureBoot\AvailableUpdates set to 0x5944 method; and it honestly went surprisingly well. That said; it's 100% thinkpads and enterprise-line Dells here; and I have the unpleasant suspicion that the 'consumer motherboard that maybe gets an update if AGESA needs to be bumped' segment has some...under-tested DBX update functionality.
We've also had a veritable torrent of WU-delivered UEFI capsule BIOS updates go out; so the OEMs seem to be doing things on their end as well.
Luckily, to the best I've been able to pin anyone down on the question, failure to update just means not being protected according to what secure boot is designed to do, rather than the system just not booting; so we shouldn't have too large an epidemic of random home wintendos just falling over and dying; and realistically home users don't exactly //
just had to check for these in powershell:
$([System.Text.Encoding]::ASCII.GetString((Get-SecureBootUEFI dbdefault).bytes) -match 'Windows UEFI CA 2023')
and
$([System.Text.Encoding]::ASCII.GetString((Get-SecureBootUEFI db).bytes) -match 'Windows UEFI CA 2023')$([System.Text.Encoding]::ASCII.GetString((Get-SecureBootUEFI db).bytes) -match 'Windows UEFI CA 2023')depending on which DB we have it stored (let me know if you know any other place). //
This script works pretty well if you don't know how to do it for yourself, or in your corporate environment. https://github.com/anomixer/Update-SecureBootCert
Why Do We Say C is Unsafe?
When the C programming language was developed circa 1970, computers had vastly smaller resources. The PDP-11, for example, came with 4KB of memory. Not gigabytes or megabytes – kilobytes. In such an environment, assembler was typically used, and C – sometimes known as “portable assembler” – is a very low-level language. With C, every bit counts, and one of its strengths is producing tight code and small executables, which is why it’s so often used for embedded systems. (Amusingly, what’s considered a “tiny embedded system” in 2026 would be a massive room-sized computer in 1970).
C is a wonderful language, but it comes with some limitations. These reflect the constraints that C was created within, but also the era. Many modern concepts simply hadn’t been invented yet.
Here are some limitations with C.
Law, not technology, is the true battlefield in the War on General Purpose Computing, a subject I've been raising the alarm about for decades now:
https://memex.craphound.com/2012/01/10/lockdown-the-coming-war-on-general-purpose-computing/
The fact that there's no technical way to enforce these restrictions means that the companies that benefit from them have to pitch their arguments to lawmakers, not customers. If you have something that works, you use it in your sales pitch, like Signal, whose actual, working security is a big part of its appeal to users.
If you have something that doesn't work, you use it in your lobbying pitch, like Apple, who justify their 30% ripoff app tax – which they can only charge because it's a felony to reverse-engineer your iPhone so you can use a different app store – by telling lawmakers that locking down their platform is essential to the security and privacy of iPhone owners:
https://pluralistic.net/2024/01/12/youre-holding-it-wrong/#if-dishwashers-were-iphones
Google lost a brutal antitrust case brought by Epic Games, makers of Fortnite:
https://pluralistic.net/2023/12/12/im-feeling-lucky/#hugger-mugger
Epic's suit contended that Google had violated antitrust law by creating exclusivity deals with carriers and device makers that locked Android users into Google's app store, which meant that Epic had to surrender 30% of its mobile earnings to Google.
Google lost that case – badly. It turns out that judges don't like it when you deliberately destroy evidence:
They say that when you find yourself in a hole, you should stop digging, but Google can't put down the shovel. After the court ordered Google to open up its app store, the company just ignored the order, which is a thing that judges hate even more than destroying evidence:
https://www.justice.gov/atr/case/epic-games-inc-v-google-llc
So it was that last month, Google found itself with just two weeks to comply with the open app store order, or else:
https://www.theverge.com/news/717440/google-epic-open-play-store-emergency-stay
Google was ordered to make it possible to install new app stores as apps, so you could go into Google Play, search for a different app store, and, with a single click, install it on your phone, and switch to getting your apps from that store, rather than Google's.
That's what's behind Google's new ban on "sideloading": this is a form of malicious compliance with the court orders stemming from its losses to Epic Games. In fact, it's not even malicious compliance – it's malicious noncompliance
The Fast Identity Online (FIDO) Alliance developed passkeys several years ago, and the technology offers numerous benefits. For example, passkeys cannot be guessed or shared. Also, passkeys resist some phishing attempts because they're unique to the sites they're created for, so they won't work on fraudulent lookalikes. Most importantly, in the age of near-constant data breaches, your passkeys cannot be stolen by hacking into a company's server or database, making the stolen data far less valuable to criminals. //
Apps or websites store your unique public key. A private key is stored on your device, in your password manager, or, if you're an Apple user, in your iCloud keychain. After your device (or iCloud) authenticates your identity, the two keys combine to grant you access to your account. //
To learn how to set up passkeys for your online accounts, check out our guide to setting up and using passkeys.
https://www.pcmag.com/how-to/no-more-passwords-how-to-set-up-apples-passkeys-for-easy-sign-ins
You know the data privacy pop-up screens? Don't immediately tap "Accept." Instead, navigate to the "Cookies" or "User Data" sections and choose the shortest available session duration. That way, your cookies will expire automatically or whenever you close your browser window. //
Because the technologies became popular around the same time, many people seem to believe that 2FA options like biometric authentication, authenticator apps, and hardware security keys are the same as passkeys.
The difference? Passkeys perform multi-factor authentication. You will log into a website using only the passkey; there is no need to enter a password and username. Depending on your privacy and security settings, the iCloud account, device, or password manager where you've stored a passkey may require you to unlock it by using your face, fingerprint, or passcode.
TL;DR: The 2021 Infrastructure Investment and Jobs Act requires all new cars sold after September 2027 to include technology that monitors whether you're impaired or distracted—and can prevent you from driving. Infrared cameras will track your eyes, breath sensors will measure alcohol, and your car can refuse to start or limit its speed. Privacy advocates warn this biometric data could be shared with insurance companies, law enforcement, or sold to data brokers.
What's coming to your car
Tucked into the 2,702-page Infrastructure Investment and Jobs Act that President Biden signed in November 2021 was a provision that few Americans noticed. Section 24220 requires NHTSA to issue safety standards mandating "advanced drunk and impaired driving prevention technology" in all new passenger vehicles.
The law gave NHTSA until November 15, 2024 to finalize rules. Enforcement begins no later than September 2027. That deadline is now 18 months away.
FreeOTP is a two-factor authentication application for systems utilizing one-time password protocols. Tokens can be added easily by scanning a QR code. If you need to generate a QR code, try our QR code generator.
FreeOTP implements open standards: HOTP and TOTP. This means that no proprietary server-side component is necessary: use any server-side component that implements these standards. We recommend FreeIPA.
On March 23, 2026, the Hong Kong government changed the implementing rules relating to the National Security Law. It is now a criminal offense to refuse to give the Hong Kong police the passwords or decryption assistance to access all personal electronic devices including cellphones and laptops. This legal change applies to everyone, including U.S. citizens, in Hong Kong, arriving or just transiting Hong Kong International Airport. In addition, the Hong Kong government also has more authority to take and keep any personal devices, as evidence, that they claim are linked to national security offenses.
The vulnerability and exploit code that exploits it were released Wednesday evening by researchers from security firm Theori, five weeks after privately disclosing it to the Linux kernel security team. The team patched the vulnerability in versions 7.0, 6.19.12, 6.18.12, 6.12.85, 6.6.137, 6.1.170, 5.15.204, and 5.10.254) but few of the Linux distributions had incorporated those fixes at the time the exploit was released.
A single script hacks all distros
The critical flaw, tracked as CVE-2026-31431 and the name CopyFail, is a local privilege escalation, a vulnerability class that allows unprivileged users to elevate themselves to administrators. CopyFail is particularly severe because it can be exploited with a single piece of exploit code—released in Wednesday’s disclosure—that works across all vulnerable distributions with no modification. With that, an attacker can, among other things, hack multi-tenant systems, break out of containers based on Kubernetes or other frameworks, and create malicious pull requests that pipe the exploit code through CI/CD work flows.
“‘Local privilege escalation’ sounds dry, so let me unpack it,” researcher Jorijn Schrijvershof wrote Thursday. “It means: an attacker who already has some way to run code on the machine, even as the most boring unprivileged user, can promote themselves to root. From there they can read every file, install backdoors, watch every process, and pivot to other systems.”
It estimates there are over 100 million consumer routers currently in active use across the US, and the FCC's order impacts the replacement cycle for every one of these devices, as new models cannot be authorized unless they secure Conditional Approval and agree to onshoring requirements.
The existing channel inventory of previously authorized router models will absorb initial demand, but that buffer is finite, and if the Conditional Approval process cannot achieve sufficient throughput within 6 to 12 months, consumers and ISPs will face a constrained selection, the GEA says.
The upshot will be that many will not be able to replace aging and outdated routers, which is more likely to leave them vulnerable to attackers taking advantage of any security flaws in them.
Firms that make router silicon such as Qualcomm, MediaTek, and Broadcom operate on global roadmaps, the report states. If the US certification pathway becomes slower or less predictable than equivalent processes in Europe or Asia, then vendors will prioritize launches in those markets, the report claims. US consumers would see delayed availability of new Wi-Fi 7 models, reduced model selection, or higher prices as companies have to cover compliance costs across fewer units sold.
Speedtest.net data reveals the most popular Wi-Fi router brands in the US, many of which could face trouble licensing new models without an FCC exemption. //
according to Ookla, TP-Link comes in second, with its Wi-Fi routers appearing in only 9.9% of speed test samples. Instead, routers from Amazon-owned Eero lead the pack, although narrowly, with a 10% share. In third is US-based Netgear at 9.6%. //
To prevent harming consumers, the FCC’s order steers clear of banning any Wi-Fi routers currently in use or sold in the US. The Trump administration will also allow vendors to apply for an exemption under the implied pretext that the company will eventually move manufacturing to the US. Whether that process favors US companies over foreign brands is a big question.
Still, as it stands, the FCC is only permitting software updates to flow to existing foreign-made Wi-Fi routers for consumers until March 1, 2027. It's a pretty ironic and alarming deadline, considering software updates keep routers safe from serious vulnerabilities.
The TP-Link WR841N router is named by the NCSC as one of the models APT28 has been exploiting, likely using CVE-2023-50224, an unauthenticated information disclosure flaw that allows an attacker to retrieve credentials through an HTTP GET request. When the threat actor has the router’s credentials, a second GET request rewrites the DHCP DNS settings, setting the primary DNS to a malicious IP and the secondary to the original primary.
The advisory lists more than 20 additional TP-Link models targeted in the campaign, //
A second cluster of attacker infrastructure received DNS requests forwarded from compromised MikroTik routers as well as TP-Link gear, and was also used in interactive operations against a smaller set of MikroTik routers "often located in Ukraine" that the NCSC said were likely of intelligence value.
Modern adversary tooling executes what security researchers call a real-time phishing relay, sometimes referred to as an adversary-in-the-middle (AiTM) attack. The mechanics are precise.
An adversary builds a reverse proxy that sits between the victim and the legitimate service. When the victim enters credentials on the spoofed page, the proxy forwards those credentials to the real site in real time.
The real site responds with an MFA challenge. The proxy forwards that challenge to the victim. The victim responds — because the page looks legitimate and the MFA prompt is real. The proxy forwards the response. The adversary receives an authenticated session.
Push notification MFA, SMS one-time codes, and TOTP authenticator apps are all vulnerable to this relay. They authenticate the exchange of a code. They do not verify that the individual completing the exchange is the authorized account holder. They cannot distinguish a direct session from a proxied one. //
The deeper problem is that the authentication architecture most organizations have deployed was not designed to answer the question that actually matters in a post-breach environment: was the authorized individual physically present and biometrically verified at the moment of authentication?
Push notifications do not answer this question. SMS codes do not answer this question. TOTP does not answer this question. USB hardware tokens answer a related but different question — they prove the registered device was present, not the authorized person. //
FIDO2/WebAuthn gets cited frequently in this conversation, and it is a meaningful step forward — but it is not sufficient on its own. Standard passkey implementations bind the credential to a device or cloud account.
Cloud-synced passkeys inherit the vulnerabilities of the cloud account: SIM swap attacks against the recovery phone number, account takeover via credential phishing, recovery flow exploitation. Device-bound passkeys prove device possession. They do not prove human presence.
Phishing-resistant authentication that closes the relay attack vector requires three properties simultaneously:
- Cryptographic origin binding: the authentication credential is mathematically tied to the exact origin domain. A spoofed site cannot produce a valid signature because the domain does not match. The attack fails before any credential is transmitted.
- Hardware-bound private keys that never leave secure hardware: the signing key cannot be exported, copied, or exfiltrated. Compromise of the endpoint does not compromise the credential.
- Live biometric verification of the authorized individual: not a stored biometric template that can be replayed, but a real-time match that confirms the authorized person is physically present at the moment of authentication.
When all three properties are present, a relay attack has no viable path. The adversary cannot produce a valid cryptographic signature from a spoofed site. They cannot relay a session because the cryptographic binding fails the moment the origin changes.
They cannot use a stolen device because the biometric verification fails without the authorized individual. They cannot social-engineer an approval because there is no approval prompt — the authentication either completes with a live biometric match at the registered hardware, or it does not complete.
AI is rapidly changing how software is written, deployed, and used. Trends point to a future where AIs can write custom software quickly and easily: “instant software.” Taken to an extreme, it might become easier for a user to have an AI write an application on demand—a spreadsheet, for example—and delete it when you’re done using it than to buy one commercially. Future systems could include a mix: both traditional long-term software and ephemeral instant software that is constantly being written, deployed, modified, and deleted.
AI is changing cybersecurity as well. In particular, AI systems are getting better at finding and patching vulnerabilities in code. This has implications for both attackers and defenders, depending on the ways this and related technologies improve.
In this essay, I want to take an optimistic view of AI’s progress, and to speculate what AI-dominated cybersecurity in an age of instant software might look like. There are a number of unknowns that will factor into how the arms race between attacker and defender might play out.