Adversaries are using AI mainly to strengthen familiar cyber and influence tactics: writing and personalizing messages, impersonating people, generating media, and analyzing large datasets. Government assessments and threat reports document these uses, but they do not show that AI has made fully autonomous hacking routine. The evidence varies by source and date, so reported activity, assessments, and forecasts need to be kept separate.
What adversaries are using AI to do
AI is being added at different points in existing workflows, from preparing a lure to analyzing information or producing material for an influence campaign. Canada’s National Cyber Threat Assessment 2025–2026 says cybercriminals and state-sponsored actors use generative and predictive AI, including large language models (LLMs), for tasks ranging from content generation to big-data analysis. The assessment says these tools are almost certainly lowering barriers to entry and improving the quality, scale, and precision of malicious activity; it does not quantify how often each use occurs. Canadian Centre for Cyber Security
- Generate or tailor content: Draft messages, articles, images, audio, or video for a target or audience.
- Impersonate: Use synthetic or altered voice and visual material to pose as a trusted person.
- Analyze information: Process bulk data, look for patterns, and inform follow-on operations.
- Work alongside conventional methods: Combine AI assistance with phishing, credential harvesting, or other established tactics.
These are complementary uses, not proof of an entirely new attack model. The practical question is what the operator is trying to accomplish and which parts AI helps with.
As an Amazon Associate I earn from qualifying purchases.
Recommended Free Tools
How AI can improve social engineering
Social engineering tries to persuade someone to disclose information, approve a fraudulent transaction, or take another action that benefits an attacker. Canada’s assessment says actors are almost certainly using LLMs to improve these attacks, including by creating convincing, personalized phishing messages and impersonating trusted people with synthetic audio or video. A fluent message or familiar-sounding voice may make a pretext more plausible; it does not by itself establish who sent it.
AI can help produce or adapt content, but the resulting message still depends on a believable scenario and a target who can be persuaded to act. Treat unexpected requests involving payment, credentials, sensitive data, or urgent action as requests to verify—especially when they arrive through a new channel or change established procedures.
#1 Best Overall
How AI appears in influence operations
Influence operations aim to shape what people believe, trust, or discuss. Canada’s threat assessment says countries including China, Russia, Iran, and Israel have reportedly incorporated deceptive AI-generated articles, images, and videos into such efforts. It also assesses that actors likely use AI to generate fictitious social-media bots and online personas. The assessment describes campaign aims such as polluting the information space, undermining trust, and sowing doubt and division; these are Canadian government assessments, not a count of verified campaigns.
Microsoft’s 2024 Digital Defense Report gives examples it attributed to state-linked activity. Its report described a May 2024 AI-generated cartoon attributed to China, June 2024 AI-generated audio of Elon Musk narrating a fabricated documentary attributed to Russia, and a video ahead of a military operation that Microsoft described as likely AI-generated and attributed to Iran. These are Microsoft’s reported findings and confidence judgments, not independent confirmation of every attribution. Microsoft Digital Defense Report 2024
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitches| Reported example | Timing and attributed actor | Evidence wording |
|---|---|---|
| AI-generated cartoon | May 2024; attributed by Microsoft to China | Microsoft report example |
| AI-generated audio of Elon Musk narrating a fabricated documentary | June 2024; attributed by Microsoft to Russia | Microsoft report example |
| Video ahead of a military operation | 2024; attributed by Microsoft to Iran | Microsoft called it “likely” AI-generated |
In its 2025 report, Microsoft described additional tactics: “AI twinning,” or digital replicas of trusted news anchors; poisoning training data; and voice cloning or masking. It also discussed AI combined with phishing, credential harvesting, and insider recruitment as a way to make operations easier to scale, more effective, and harder to trace. These are Microsoft’s reported trends and analysis, not a global prevalence measurement. Microsoft Digital Defense Report 2025
What the evidence says about scale and capability
The U.S. Government Accountability Office’s 2026 Science & Tech Spotlight says many malicious uses of generative AI already exist and tactics are evolving. It also relays an academic study’s estimate that generative AI could reduce malicious users’ phishing-attack costs by more than 95 percent. That is an estimate reported by GAO, not a GAO-run experiment or a measured cost reduction that applies to every real-world phishing campaign; the GAO page does not provide the underlying study’s methods. U.S. Government Accountability Office
Rank #3
There is also evidence against treating every AI-related threat as a mature or autonomous capability. CrowdStrike’s 2025 Threat Hunting Report characterized much adversarial generative-AI use as iterative and experimental while discussing examples in cyber and influence operations. That describes CrowdStrike’s own reporting and visibility, not a census of all activity. CrowdStrike 2025 Threat Hunting Report
A UK government assessment published in 2023 judged that fully automating computer hacking by 2025 was unlikely, and that criminals were less likely to successfully create novel malware with generative AI by then. Those were forecasts made for a horizon through 2025—not a finding about the present state of every threat actor in 2026. The assessment also excluded military risks. UK Department for Science, Innovation and Technology assessment
How to interpret the dated warnings
The UK assessment’s central judgment was: “Over the next 18 months, generative AI (artificial intelligence) is more likely to amplify existing risks than create new ones.” Its publication-era outlook extended through 2025; it should not be presented as a new 2026 prediction. The same assessment said digital risks were the most likely to manifest with high impact within its horizon, while also warning about risks to political systems and society and risks as AI becomes embedded in physical systems. UK government assessment, October 2023
Other assessments have different time frames. The U.S. Department of Homeland Security’s Homeland Threat Assessment 2025, published in September 2024, expected malicious use of generative AI to continue in cyber activity and malign influence during 2025, citing lower technical thresholds and improved personalization and scale. That, too, was a forecast for a specified period rather than a retrospective measurement of all activity. DHS Homeland Threat Assessment 2025
Best Value
Read reports with their scope in mind: a government assessment expresses an agency judgment, while a security vendor report reflects that company’s telemetry and analytic framework. Neither alone establishes a complete global count. A forecast describes what its authors expected at the time, not what later evidence necessarily confirmed.
How can malicious efforts be countered?
No single safeguard addresses every use of AI. Responses work best when they combine technical controls with verification procedures and clear ways to report suspicious activity.
Reduce the chance that a convincing message triggers an action
- Verify unusual payment, credential, or data requests through a known, independent channel—not by replying to the message or calling a number it supplies.
- Require an established approval process for sensitive transactions and changes to account or payment details.
- Train staff to report suspicious messages and impersonation attempts, including plausible messages that do not contain obvious spelling or formatting errors.
Protect systems and AI services
- Use layered cyber defenses appropriate to the organization, including controls that address phishing and credential theft.
- Assess AI systems’ attack surfaces and keep safeguards under review as systems and manipulation techniques change.
- For AI systems, consider safeguards such as reinforcing protections through human feedback, filtering user instructions, and using separate generative-AI technology to detect malicious attacks. GAO notes these as mitigations identified by NIST, while also warning that techniques for manipulating systems evolve and security work must continue. GAO Science & Tech Spotlight
Handle suspicious media carefully
- Do not treat a familiar face, voice, or polished presentation as proof of authenticity.
- Check consequential claims against reliable sources and seek confirmation from the person or organization through a trusted channel.
- Preserve and report suspicious messages or media through the relevant workplace, platform, or institutional reporting process instead of amplifying unverified material.
These measures reduce opportunities for deception and improve detection; none guarantees that an attack or misleading content will be identified in advance.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
