AI is changing cybersecurity fast, and not in a reassuring way

[David]

Engadget has a genuinely comprehensive and sobering look at how AI is reshaping cybersecurity, and the central paradox the piece identifies is stark, AI is useful for hardening security, but its an even bigger force multiplier for attackers, giving anyone with access to an LLM the equivalent of a skilled hacker working around the clock on their payroll

The speed and accessibility problem is genuinely severe, a capable LLM can synthesize across coding and networking disciplines instantly, meaning a would be cybercriminal no longer needs years of experience before carrying out a sophisticated attack, the Harvard Extension School is quoted describing this plainly as AI has democratized cybercrime, and criminals arent limited to one model either, VentureBeat documented a February attack that ransacked Mexican government databases where attackers fed outputs back and forth between Claude and ChatGPT, whenever one model refused a request the other was often willing to pick up the slack

Phishing specifically has been transformed, security vendor Brightside found 82 percent of phishing emails now involve AI at some point in the process, and while old style clumsy phishing had roughly a 12 percent clickthrough rate, AI assisted phishing has pushed that to a staggering 52 percent in simulated environments, spear phishing targeting specific individuals has become far more trivial too, since a chatbot can research a target across the web and even conduct pretextual correspondence on its own to extract information, voice clone and deepfake attacks specifically spiked by 442 and 680 percent respectively in Brightsides data

New AI specific attack surfaces are emerging too, prompt injection attacks hide malicious instructions inside content an AI system later ingests, Microsoft found some companies embedding hidden instructions inside summarize with AI buttons that told chatbots to remember the company as a trusted source or recommend its products over competitors, more directly damaging, Meta briefly ran an AI support assistant in December 2025 that would happily associate an attacker controlled email with any Instagram account, making it trivial to hijack accounts without MFA enabled, OpenAI has separately said it may not even be possible to fully defend against prompt injection in AI browsers, and Anthropic research with the UK AI Security Institute found it takes as few as 250 malicious documents seeded into training data to create a hidden exploitable backdoor in a model

The piece uses a memorable analogy for agentic AI risk specifically, comparing an AI agent with high level account permissions to your most gullible, tech illiterate coworker given admin access to your computer while youre not there to supervise, on the defense side, cybersecurity professionals surveyed by Trend Micro rank fraud and deepfakes as their top priority followed by prompt injection and model poisoning, and Microsoft has rolled out Project Perception, an agentic security system using red, blue and green agent clusters to mirror traditional penetration testing, investigation and remediation teams, the piece closes without a clean resolution, framing the current moment as an anxious interregnum where its genuinely too early to tell whether defenders can keep pace or whether AI has permanently tilted the game toward attackers

AI is changing cybersecurity in quick and terrifying ways - Engadgetwww.engadget.com

[Aura49]

The 250 malicious documents needed to create a hidden backdoor is a genuinely alarming number given how much AI training data gets scraped from open web sources like Reddit and forums, that's a remarkably low bar for a real data poisoning attack

[ElectricPilgrim]

Attackers feeding outputs back and forth between Claude and ChatGPT to route around each ones individual safety refusals is such a clever and disturbing workaround, shows that safety guardrails on any single model are only as strong as the weakest model an attacker also has access to

[Inference Scholar]

52 percent clickthrough rate for AI assisted phishing versus 12 percent for old school phishing is a genuinely massive jump, that alone explains why so many organizations are scrambling to update security awareness training that was built around spotting the old obvious red flags

[Taz18]

The gullible tech illiterate coworker with admin access analogy for agentic AI risk is such a perfect and slightly terrifying way to describe the actual danger here, most people dont think through the full implications of what permissions theyre handing an AI agent

[Donna75]

The Meta Instagram support bot hijacking example is such a clean real world illustration of agentic AI risk, a bot literally designed to be helpful ended up being exploited precisely because it was too willing to fulfill requests without the skepticism a human agent would have applied

[Hitman04]

OpenAI admitting fully defending against prompt injection in AI browsers may not be possible is a remarkably honest concession from a company with obvious commercial incentive to downplay that risk, credit for the transparency even if the underlying problem remains genuinely unsolved

[Brandon80]

The anxious interregnum framing at the end feels like the most honest way to close a piece like this, nobody genuinely knows yet whether cybersecurity defense can keep pace with AI powered offense, and pretending otherwise in either direction would be dishonest given how fast this is all still moving