TLDR: AI political astroturfing in Kenya is one of the key findings in Anthropic’s latest threat intelligence report, Detecting and countering misuse of AI: September 2026. The report details how malicious actors attempted to use Claude for influence operations, cyberattacks, surveillance, scams and other harmful activities.
Anthropic has released a new threat intelligence report detailing how actors around the world have attempted to misuse its Claude AI models for cyber operations, political influence campaigns, surveillance, scams and other harmful activities.
The report, Detecting and countering misuse of AI: September 2026, was published on September 10, 2026. It covers operations that Anthropic says its Threat Intelligence team identified and disrupted between December 2025 and August 2026, an eight-month period.
The 154-page report covers seven areas: cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development and illicit model distillation. Anthropic also makes an important qualification: the cases are not examples of typical Claude use, but some of the most notable and novel cases of misuse it has identified.
Among the countries featured is Kenya, where Anthropic says it disrupted an operation using Claude to generate political social media content designed to look like genuine grassroots public opinion.
AI Political Astroturfing in Kenya: What Anthropic Found
The Kenyan case is identified as GTG-54004: Disrupting a domestic coordinated inauthentic behavior campaign in Kenya.
GTG stands for Generative Threat Group, Anthropic’s internal designation for actors it observes abusing AI systems.
According to Anthropic, it identified and removed an account used by a single actor to mass-produce Kenyan political content. The actor used Claude over several sessions to generate batches of exactly 50 tweets, while specifically asking the model to make them look like spontaneous grassroots commentary rather than a coordinated campaign.
In simple terms, the aim was to make organised messaging look as though many ordinary people had independently arrived at and were sharing the same opinion.
This type of activity is commonly known as astroturfing: creating the appearance of grassroots public support or opposition when the activity is actually coordinated.
Which political narratives were being pushed?
Anthropic says a large portion of the AI-generated posts praised Hon. James Opiyo Wandayi, EGH, is the Cabinet Secretary for Energy and Petroleum, in relation to stopping a scheduled Kenya Power electricity tariff increase.
The posts used hashtags including #PowerReliefKE and #PoweringTheNewKenya.
The same network also produced content claiming that Kenya’s United Opposition coalition was falling apart ahead of the 2027 election, with posts directly targeting Rigathi Gachagua and former President Uhuru Kenyatta.
Anthropic describes the operation as using one playbook to amplify both pro-government and anti-opposition narratives, which it says was consistent with a coordinated electoral communications effort.
There is, however, an important distinction.
Anthropic says it found no evidence of government involvement. It assessed the activity as an entirely domestic Kenyan operation.
The company says the campaign’s pro-administration tone and use of particular hashtags suggested that it was plausibly aligned with the ruling coalition, but Anthropic says it did not identify the specific organisation responsible.
That distinction matters. The report does not establish that the Kenyan government, a political party or any named politician commissioned the campaign.
How was Claude being used?
What stands out in the Kenya case is not that AI came up with the political strategy.
Anthropic says the actors had already decided the ideological messages before using Claude.
They supplied the topics, talking points and hashtags, then used the model to turn them into batches of 50 themed posts formatted for distribution. Claude was also used to humanise and refine pre-written material, helping give the posts greater variety and the appearance of authenticity.
In several sessions, Anthropic says the output was even packaged into an interactive copy-all widget, ready for deployment.
Therefore, the Kenya case shows how generative AI can be used to scale a political communications operation. Instead of manually writing dozens of different tweets around one talking point, an operator can use AI to produce variations quickly while trying to make them appear as independent voices.
For communicators and social media professionals, that is an important distinction.
The concern is not simply whether somebody used Generative AI to help write a post but the use of AI to make a coordinated campaign appear to represent independent public opinion.
The same workflow was used to market Kenyan retail brands
Anthropic also found that the actor behind the political activity used the same AI workflow for commercial marketing.
The report says the workflow was used for Kenyan retail brands under a marketing persona called “SHANKI” or “Elkins Marketer.”
Anthropic says one retail promotional broadcast was repackaged as apparently organic tweets, with links inserted into every fifth post.
This part of the report should interest marketers and agencies.
The same technology and workflow that can generate dozens of variations of a commercial campaign can also be applied to politics. The difference is not necessarily the tool. It is the purpose, transparency and way the campaign is presented to the public.
Anthropic describes this as fitting a pattern it has observed in Kenya involving agencies paying local influencers to manipulate narratives.
Did the campaign actually influence Kenyans?
According to Anthropic, there is no evidence that this particular operation achieved significant reach.
Anthropic assessed the Kenyan network as Category One on the Breakout Scale, the lowest category used in the report to measure the reach of influence operations.
The company says the activity remained confined to its network of fake accounts on one platform and did not break out to a genuine audience.
This is important because the existence of an influence operation does not automatically mean it was successful.
Anthropic says this is a wider pattern across several of the campaigns it investigated. Because AI companies can sometimes detect an operation while content is still being produced, some campaigns can be disrupted before they build a real audience. The report says much of the influence content it found generated little or no authentic engagement.
Figure 17 in the report shows examples of inauthentic commenting accounts amplifying content relating to Wandayi and the electricity tariff issue.
OpenAI helped flag the Kenyan activity
Another notable detail is how Anthropic says the investigation began.
According to the report, OpenAI shared a tip about recidivist activity on its platform. Anthropic then conducted its own investigation into suspected coordinated inauthentic behaviour in Kenya.
Anthropic says it subsequently removed the account and organisation behind the activity and developed detections based on the operation’s behavioural patterns.
The case highlights the importance of cooperation between AI companies when the same actors may be using multiple AI services as part of one operation.
Kenya is not the only African country in the report
Anthropic’s report documents several other operations involving African countries. The cases differ considerably, ranging from political influence campaigns to surveillance.
Central African Republic: Russian state-aligned influence operation
In the Central African Republic (CAR), Anthropic says it disrupted an operation run by a Russian-speaking actor in Bangui that formed part of a Russian state-aligned foreign information manipulation campaign.
The operation used Claude to help produce material distributed through Radio Lengo Songo, while coordinating with Russian state media organisations including RT, Sputnik Afrique and TASS.
Anthropic says much of the sampled messaging was pro-CAR government, pro-Wagner, anti-France and anti-CAR opposition.
The use of Claude went beyond writing social media posts.
According to the report, the model was used to help create contracts, job descriptions, staff-scoring systems and political content. The operation also tracked opposition political figures, drafted talking points and produced forged CAR government documents.
Unlike the Kenyan case, Anthropic assessed the CAR operation as Category Four because the material reached audiences through FM radio, Telegram and local news outlets.
Democratic Republic of Congo: fake news websites and X accounts
The report also describes a commercial influence-as-a-service operation that targeted audiences across six continents.
The network included approximately 70 fabricated news websites, 70 linked X accounts and more than 250 inauthentic commenting accounts. Anthropic traced the operation to a France-based digital advertising agency.
The Democratic Republic of Congo was one of the operation’s major areas of focus.
Anthropic identified 318 articles across the fake news network focusing on the DRC, including narratives around regional mineral agreements and tensions with Rwanda.
The company says it found signals suggesting that one or more customers may have had an interest in the DRC-Rwanda conflict, but it could not independently identify who commissioned the material and found no evidence of government direction.
Sudan: influence operations hiding their origin
Another case in the report involved the Sudan conflict.
Anthropic says it disrupted a UAE-directed influence operation that used approximately 300 inauthentic influencer accounts. The operation also created a front NGO using the identity of a real organisation, profiled members of the European Parliament and journalists, and compiled dossiers on UN Special Rapporteurs.
Anthropic says the actor also ghost-wrote testimony intended for the UN Human Rights Council while taking steps to prevent the material from revealing its connection to the UAE.
The report says the identity of a real Sudanese human-rights organisation was borrowed so that material supporting one side of the Sudan conflict could appear to come from independent local witnesses.
Mali: Claude used to help engineer a surveillance platform
The report’s African cases are not limited to political influence.
In Mali, Anthropic says a single Claude subscriber, believed to be a Bamako-based independent consultant working with the country’s state intelligence service, used Claude as engineering support while developing a surveillance platform known as Lakana 360.
Anthropic says the system was designed to monitor roughly 25 million SIM cards across all three of Mali’s national mobile operators.
According to the report, the system included capabilities involving call records, SMS and voice traffic, cross-SIM voice identification, watchlists, VPN and encryption-user flagging, and matching individuals against government registries.
Anthropic banned the account involved. However, it says the surveillance platform had been deployed locally using an on-premises model, meaning its enforcement action could disrupt further development using Claude but could not remove the system that had already been deployed.
What else does the Anthropic report reveal about AI misuse?
The report presents a much wider picture of how AI tools can be incorporated into harmful operations.
Within influence campaigns, Anthropic found actors using Claude as a newsdesk, helping smaller teams produce material at a scale that would previously have required more staff.
It also documented the use of AI to create fake personas, rewrite or localise content, conceal the original source of political narratives, produce internal operational documents and make coordinated content appear more organic.
Beyond influence operations, the report covers cyberattacks, surveillance, fraud, weapons development and other forms of misuse.
The broader concern is therefore not simply that AI can generate misleading text. AI can reduce the amount of time, staffing and technical expertise required to operate an existing campaign at scale.
What does this mean as Kenya approaches the 2027 election?
The Kenya operation documented by Anthropic did not achieve significant reach, according to the report.
But what it demonstrates deserves attention.
A single actor was able to provide political talking points and use AI to turn them into dozens of different posts designed to resemble independent public commentary.
As Kenya moves towards the 2027 election, this raises questions that journalists, marketers, communication professionals, researchers, political actors and social media platforms will need to consider.
How do we distinguish genuine public opinion from coordinated amplification?
A trending hashtag or thousands of posts around one issue can look like strong public sentiment. But volume alone does not tell us whether the conversation developed organically.
What happens when paid influencers, fake accounts, genuine citizens and AI-assisted content all participate in the same conversation?
Separating those groups may become increasingly difficult when AI can make coordinated posts sound different from one another.
Should political digital campaigns be more transparent about who paid for or coordinated them?
The Anthropic case demonstrates why identifying the source of online political narratives is becoming as important as analysing the content itself.
Are social media platforms prepared to detect coordination rather than simply AI-generated text?
Detecting whether a paragraph was written with AI may become less useful than looking for patterns across accounts: timing, repeated talking points, shared hashtags, unusual amplification and networks repeatedly interacting with one another.
For those of us working in social listening, media monitoring and digital communication, there is another important question.
What does a spike in mentions actually represent?
If hundreds of accounts suddenly begin praising or attacking the same person, organisation or policy, analysing volume and sentiment alone may give a misleading picture. Increasingly, monitoring needs to consider how a narrative is spreading, which accounts are driving it, whether the amplification appears coordinated and whether what looks like public sentiment is genuinely organic.
That is an analytical implication of Anthropic’s findings, rather than a claim made directly by the report.
The issue is bigger than AI-generated content
It would be easy to read the Kenya case simply as another example of people using AI to write political posts.
The report points to a bigger issue.
In the case Anthropic documented, humans selected the political positions, talking points and hashtags. AI was then used to increase the volume and make the material appear more natural.
That makes the issue less about whether content is “AI-generated” and more about how AI can help organised actors manufacture the appearance of authentic public participation.
For marketers and communication professionals, the ethical question is particularly relevant. The same tools that make it easier to create campaigns, adapt content and reach audiences can also be used to disguise coordinated messaging as grassroots conversation.
As the 2027 election approaches, digital literacy will therefore need to extend beyond spotting deepfakes and obviously AI-generated images.
People may increasingly need to ask:
Who is behind this message? Why are so many accounts saying something similar? Is this really an organic conversation? And who benefits if I believe that it is?
Those questions may become just as important as asking whether a piece of content was created using AI.
Read the Anthropic report
This article is based on Anthropic’s Detecting and countering misuse of AI: September 2026, published on September 10, 2026. The report covers malicious activity Anthropic says it identified and disrupted between December 2025 and August 2026.
The full report and PDF are available directly from Anthropic
Read and download Detecting and countering misuse of AI: September 2026
