How AI Is Weaponizing Information: The New Era of Propaganda, Surveillance, and Political Manipulation

5 days ago

9 min read

SWE-2 and the New Economics of AI Coding Agents

Artificial intelligence coding agents are entering a new phase in which raw model intelligence is no longer the only metric that matters. For developers and businesses, the more important question is increasingly whether an AI system can solve difficult engineering problems accurately while minimizing unnecessary computation, latency, tool usage and cost.

Cognition’s SWE-2 illustrates this shift. Built by the company behind Devin, SWE-2 is a reinforcement-learning post-trained coding model based on Moonshot AI’s 2.8 trillion-parameter Kimi K3. Cognition reports that SWE-2 reaches 50.0% on FrontierCode 1.1 Main, placing it within one percentage point of Fable 5.1 while operating at 64% lower cost. More importantly, the model introduces a training strategy designed to improve the entire relationship between coding capability and computational expense rather than optimizing intelligence in isolation.

The result is a notable development in the economics of AI-assisted software engineering. SWE-2 demonstrates how reinforcement learning, cost-aware objectives, efficient inference and better verification can combine to produce coding agents that are not merely more capable, but more economical to operate.

SWE-2 Changes the Optimization Target for Coding Models

Traditional model development often emphasizes benchmark accuracy or solve rate. Those measurements remain important, but coding agents introduce an additional variable: the amount of computation required to reach an answer.

An agent that solves a programming problem but spends excessive time exploring irrelevant files, generating unnecessary tokens or repeatedly testing the same hypothesis may be less useful than a slightly less capable system that reaches the correct solution efficiently.

Cognition therefore frames SWE-2 around the cost-performance Pareto frontier. In simple terms, the objective is to maximize coding performance at different levels of computational expenditure.

This produces a more practical model of AI performance:

Dimension Conventional optimization SWE-2 approach
Primary objective Capability Capability plus cost
Reasoning Often optimized independently Multiple effort levels trained together
Compute Secondary consideration Explicitly incorporated into reward
Exploration Thoroughness can dominate Focused exploration
Evaluation Solve rate Solve rate, cost and behavior
Deployment Model quality Model quality plus operating economics

The significance extends beyond one model. As AI coding agents become part of professional software development, inference economics can influence whether organizations deploy them continuously or reserve them for high-value tasks.

A 2.8 Trillion-Parameter Foundation With Additional RL Gains

SWE-2 is post-trained from Kimi K3, a 2.8 trillion-parameter model that had already undergone substantial reinforcement learning for agentic coding.

Rather than assuming that a powerful base model has exhausted its potential, Cognition reports that additional reinforcement learning produces another 5 to 6 points of improvement across many benchmarks.

That finding highlights the growing importance of post-training.

A large foundation model can possess broad programming knowledge, but autonomous software engineering requires more specialized behaviors. An agent must interpret a repository, identify relevant files, formulate a plan, modify code, run tests, diagnose failures and verify that its solution actually satisfies the task.

Reinforcement learning can optimize these behaviors around concrete outcomes rather than merely predicting plausible programming text.

The SWE-2 approach consequently reflects a broader transition in AI development: scaling the underlying model remains important, but increasingly sophisticated post-training can determine how effectively that intelligence translates into real-world work.

The Benchmark Picture Reveals Both Strengths and Weaknesses

Cognition reports strong results across several coding evaluations.

Benchmark SWE-2 Kimi K3 Fable 5.1 GPT-5.6 Sol GPT-6 Astra
FrontierCode 1.1 Main 50.0% 44.2% 50.9% 47.5% 53.3%
DeepSWE 1.1 73.0% 68.5% 67.4% 72.7% 74.1%
Terminal-Bench 2.1 92.8% 88.3% 91.4% 88.8% 89.9%
Terminal-Bench 4 27.3% 21.5% 55.8% 37.3% 57.9%

The numbers demonstrate both the progress and the remaining challenges.

SWE-2 leads the listed models on Terminal-Bench 2.1 and substantially improves over its Kimi K3 foundation across every benchmark shown. Its 50.0% FrontierCode score also places it close to Fable 5.1.

However, Terminal-Bench 4 exposes a significant weakness. SWE-2's 27.3% result is far below Fable 5.1 and GPT-6 Astra. This matters because benchmark specialization can conceal important differences in agent behavior.

The lesson is that no single benchmark should define coding-agent intelligence. Real-world evaluation needs to consider repository complexity, tool interaction, reliability, verification, cost and the ability to generalize across unfamiliar engineering environments.

Focused Exploration May Be More Valuable Than Maximum Exploration

One of SWE-2's most interesting reported improvements is behavioral rather than purely numerical.

Cognition says SWE-1.7 often explored repositories extensively before making changes. That caution could improve reliability on difficult problems, but it also created inefficiency on straightforward tasks.

SWE-2 attempts to replace broad exploration with focused exploration.

On FrontierCode 1.1 Main, Cognition reports that SWE-2 medium achieves a higher score than SWE-1.7 while using 58% fewer turns and costing 81% less on average. Mean steps per run fall from 127 for SWE-1.7 to 53 for SWE-2 medium, 80 for high and 98 for max.

The model also reaches its first meaningful code edit after a median of 18 steps, compared with 48 steps for SWE-1.7.

This represents an important conceptual change in autonomous coding. Intelligence is not simply the ability to reason longer. It is also the ability to determine when additional reasoning is unnecessary.

For an AI coding agent, recognizing which files matter can be as important as understanding the programming language itself.

Three Behaviors That Matter in Real Software Engineering

Cognition identifies stronger test coverage, resourcefulness and verification discipline as notable behavioral improvements.

Better test coverage can allow an agent to detect regressions and edge cases rather than stopping once an implementation appears superficially correct.

Resourcefulness allows the model to pursue alternative routes when its preferred tool or integration is unavailable. In one reported case, SWE-2 reconstructed information from accessible Slack history when an MCP integration was unavailable.

Verification discipline may be even more consequential. Instead of automatically accepting a user's hypothesis, the model can reassess the claim and gather evidence by executing artifacts or checking results.

These behaviors point toward a more mature definition of AI coding intelligence.

A useful coding agent should not merely produce code. It should operate as an engineering system capable of investigation, implementation and validation.

Training the Entire Cost-Performance Frontier

The technical centerpiece of SWE-2 is Cognition's cost-aware reinforcement-learning methodology.

The reward function is represented as:

R = S − λₑC

where S represents successful completion, C combines rollout cost and time, and λₑ determines the cost penalty for a particular effort level.

The important idea is how Cognition selects λ.

Instead of treating the penalty as an arbitrary hyperparameter, Cognition relates it to the local slope of the base model's cost-performance Pareto frontier. The resulting reward objective is designed so that its iso-reward line is tangent to the existing frontier.

This geometric formulation has a practical purpose. If the cost penalty is poorly selected, reinforcement learning might reward the model for becoming cheaper without actually improving the useful frontier. A slope-matched penalty instead encourages improvements that move the tradeoff between cost and solve rate in a favorable direction.

The approach also enables multiple reasoning-effort levels to be trained during a single reinforcement-learning run.

That is significant because users do not have identical coding problems. A small bug fix should not require the same computational budget as a complex architectural change.

Efficient RL Requires Efficient Infrastructure

Training a large coding model through reinforcement learning creates a substantial infrastructure challenge. Rollouts are expensive because the model must repeatedly interact with coding environments while generating long sequences.

Cognition reports several engineering improvements.

A prefill delayer groups nearby requests within the GPU scheduler, increasing tokens per minute per GPU and tokens per second per request by 10% to 20%. The company accepts a higher time to first token in exchange for improved overall throughput.

SWE-2 also uses DSpark speculative decoding. In this architecture, a smaller draft model proposes tokens while the main policy verifies them. Because the policy changes throughout reinforcement learning, the acceptance rate of speculative tokens can decline.

Cognition addressed this by training a new draft model with SpecForge and then continuing to train the draft model online as the main policy evolved. The resulting system reportedly achieved 15% longer accepted sequences.

Low-precision computation also plays an important role. NVFP4 and FP8 kernels, together with quantization-aware training, reduce memory requirements while helping keep inference behavior closer to the training environment.

These optimizations demonstrate an increasingly important reality of advanced AI: algorithmic improvements and systems engineering are becoming inseparable.

Better Data and Stronger Verifiers

Model intelligence ultimately depends on the quality of the environments used for training.

Cognition reports that it tripled the number of reinforcement-learning environments for SWE-2 and expanded the diversity of repositories represented in its training data.

The company also introduced instruction-following overlays, allowing the model to practice maintaining multiple requirements while still solving the underlying engineering problem.

Perhaps most interesting is the use of a recursive verifier-improvement process.

As stronger models discover ways to exploit weaknesses in evaluation systems, those weaknesses can produce false positives or false negatives. Cognition used earlier SWE-2 checkpoints to identify such problems and refine the verifiers.

This creates a feedback loop:

stronger model → harder-to-detect failure modes → improved verifiers → better training signals → stronger model.

For agentic AI, this type of feedback loop could become as important as increasing parameter counts.

What SWE-2 Means for Businesses

The commercial significance of SWE-2 extends beyond benchmark rankings.

Lower inference cost can make continuous AI-assisted development economically viable. Instead of using expensive reasoning models only for exceptional tasks, organizations may be able to deploy different effort levels according to task complexity.

A lightweight effort mode could handle routine maintenance, documentation or straightforward bug fixes. Higher effort could be reserved for architectural changes, complex debugging or security-sensitive work.

This resembles how organizations already allocate human expertise. Not every software problem requires a senior architect, and AI systems may increasingly adopt similar resource allocation strategies.

However, the absence of open weights and a standalone API limits SWE-2's flexibility. According to the supplied material, it operates through Devin, including Desktop and CLI, with Web and Fusion availability being rolled out.

That makes SWE-2 primarily a component of an integrated coding-agent ecosystem rather than a general-purpose model that organizations can independently deploy on their own infrastructure.

The Future of Coding AI Is About Efficiency as Much as Intelligence

SWE-2 reflects a broader change in how AI coding systems are being engineered.

The next generation of coding agents will likely compete across several dimensions simultaneously:

Coding accuracy
Reasoning efficiency
Inference cost
Tool-use reliability
Repository understanding
Test quality
Verification behavior
Security and trustworthiness
Latency
Generalization

A model that dominates one dimension while performing poorly across the others may not become the most useful engineering system.

The SWE-2 results suggest that the strongest systems may instead be those capable of dynamically deciding how much computational effort a task deserves.

That capability has implications for the economics of software development. If AI agents can complete more engineering work while consuming fewer computational resources, organizations can expand their use without proportionally increasing inference expenditure.

Conclusion

SWE-2 represents an important step in the evolution of AI coding agents because it treats capability and efficiency as a connected optimization problem.

Its reported 50.0% FrontierCode 1.1 Main score, 92.8% Terminal-Bench 2.1 result and substantial reduction in turns and cost against SWE-1.7 demonstrate the potential of combining large-scale reinforcement learning with cost-aware training and optimized inference infrastructure.

Yet its weaker Terminal-Bench 4 performance also provides an important reminder: progress in autonomous software engineering cannot be measured through a single score.

The larger story is the shift toward AI systems that know not only how to solve a problem, but how much effort the problem deserves.

Artificial intelligence is changing the economics of influence operations. What once required teams of writers, analysts, translators, software engineers, social media operators, and intelligence researchers can increasingly be assembled around a small number of people using general-purpose AI systems.

Recent investigations into malicious uses of AI illustrate a significant evolution. The central development is not simply that AI can generate propaganda or misinformation. It is that AI can become an operational layer connecting research, targeting, content production, automation, persona management, synthetic media, surveillance, and distribution. This turns influence operations from relatively labor-intensive campaigns into scalable systems capable of adapting messages to different audiences and political environments.

The emerging threat therefore extends beyond fake news. AI is becoming infrastructure for information warfare.

From Content Generation to End-to-End Influence Systems

Earlier generations of online influence operations typically depended on human operators performing separate tasks. Researchers gathered information, writers produced narratives, translators localized them, designers created visual material, and social media teams distributed the finished product.

AI can collapse many of these functions into one workflow.

Investigations involving Claude demonstrate several recurring applications. Operators used AI to generate original articles, rewrite legitimate reporting, translate material, modify political framing, create fake personas, develop social media posts, analyze large datasets, construct targeting profiles, and prepare material for automated publication.

The important distinction is between AI as a writing assistant and AI as an operational platform.

A writing assistant improves productivity for one task. An operational platform can coordinate a sequence of tasks, preserve instructions, process large quantities of information, and produce standardized outputs that flow directly into automated systems.

This distinction explains why relatively small groups can now attempt operations that previously required considerably larger organizations.

The Industrialization of Synthetic News

One of the clearest developments is the emergence of synthetic news ecosystems.

In one investigated network, roughly 70 apparently independent news properties were connected through shared infrastructure. The operation reportedly produced at least 8,913 articles in approximately 20 languages. Rather than simply generating random propaganda, the system was engineered around repeatable publishing requirements, including structured outputs, formatted HTML, character constraints, and internal linking.

That architecture reveals an important objective: scale combined with the appearance of legitimacy.

Fake outlets can be given different names, geographic identities, editorial styles, and fictional journalists while operating from a common technological foundation. AI makes it possible to populate those outlets continuously and customize their content for individual markets.

More sophisticated operations can also take a legitimate article and transform it into multiple politically divergent versions. A single source story can therefore become several narratives, each optimized for a different ideological or national audience.

This creates a new problem for information integrity. The question is no longer simply whether a story is fabricated. Investigators must determine whether an authentic source has been selectively transformed, stripped of context, politically reframed, or redistributed through an artificial network.

Narrative Laundering and the Illusion of Independent Confirmation

A particularly powerful technique is narrative laundering.

Information laundering occurs when a claim is moved through several apparently independent sources until its original provenance becomes difficult to recognize. AI accelerates this process by allowing the same underlying narrative to be rewritten repeatedly, translated into multiple languages, and adapted to different editorial voices.

The resulting ecosystem can manufacture the appearance of corroboration.

For example, a claim originating with a politically motivated source can be rewritten as a local news story, republished by another website, summarized on social media, and subsequently cited as evidence by another account. Each stage appears separate even though the underlying narrative originated from the same source.

AI dramatically lowers the cost of producing these variations.

The danger is especially significant for search engines and recommendation systems. Large volumes of semantically related pages can create an artificial information environment in which repetition is mistaken for independent confirmation. Search-engine optimization can therefore become part of an influence operation rather than merely a commercial marketing technique.

Microtargeting Meets Generative AI

The next stage is the combination of generative AI with detailed demographic and behavioral data.

An investigated election-manipulation platform targeting Malaysia reportedly processed census and electoral information at constituency level and used that information to organize political targeting across all 222 parliamentary constituencies. The system focused on sensitive fault lines involving race, religion, and royalty while managing a large network of artificial social media accounts.

This represents a fundamental change in political persuasion.

Traditional propaganda generally broadcasts one message to a large population. AI-enabled influence systems can generate many versions of the same strategic narrative, each designed for a specific audience segment.

The technical workflow can be understood as:

Collect data about audiences, communities, interests, and political environments.
Segment targets according to demographic or behavioral characteristics.
Generate tailored narratives using language and cultural context appropriate to each segment.
Create multiple content formats, including articles, posts, images, scripts, and videos.
Distribute through coordinated accounts designed to create artificial engagement.
Measure responses and adjust subsequent messaging.

This resembles modern digital advertising infrastructure, except the objective may be political manipulation rather than legitimate persuasion.

Automation Changes the Scale of Political Deception

Automation is perhaps the most consequential element.

A Bangladesh-based operation reportedly rotated through 29 Claude accounts and used custom software to generate standardized batches containing fabricated headlines, detailed false stories, and image-generation prompts. The material was then moved through cloud storage and converted into multimedia content for scheduled publication.

The significance lies in the removal of continuous human involvement.

Once an automated pipeline is established, humans can concentrate on selecting objectives and strategic narratives while software handles repetitive execution. AI becomes the production engine, while automation becomes the distribution mechanism.

This creates a new asymmetry. A campaign does not necessarily need thousands of employees to produce thousands of pieces of content. A small number of operators can instead construct systems that repeatedly execute predefined instructions.

The same principle applies beyond propaganda. Investigations have documented AI-assisted surveillance platforms, recruitment systems, intelligence analysis, and software development, showing that the technology can increase operational capacity across the broader security ecosystem.

AI and State-Sponsored Information Warfare

Another important development is the integration of AI into established state-media and propaganda structures.

Investigations found instances where Claude-generated material entered existing Russian state-aligned media pipelines. AI was used to transform source material into articles, localized Spanish-language content, social media posts, broadcast tickers, captions, and voiceover scripts.

This matters because established distribution channels already possess authentic audiences.

A fabricated website may struggle to attract genuine readers. A state media organization with an established audience does not face the same limitation. AI can therefore function as a productivity multiplier inside an existing propaganda apparatus.

Similar patterns appeared in investigations involving Iranian state-aligned institutions, where AI was reportedly used to develop campaign doctrine, persona systems, targeting databases, organizational plans, multilingual content, and attribution-laundering strategies.

The strategic value is not merely faster writing. AI can help transform broad political objectives into operational documents and repeatable procedures.

Surveillance Is Becoming an AI Engineering Problem

The same transformation is occurring in surveillance.

Investigated operations included attempts to use AI to analyze large collections of social media material, identify potential targets, generate profiles, score political sensitivity, and support recruitment or monitoring.

In one case involving Iranian audiences, an AI system was reportedly used to analyze tens of thousands of archived messages and construct detailed profiles of individuals. Another operation involving Uyghur targets demonstrated how an AI assistant could support multilingual outreach, translation, conversational role-playing, and information collection.

The significance is profound because surveillance traditionally depends on specialized analysts.

Large datasets are difficult for humans to process manually. AI can rapidly transform unstructured information into structured records, classify material, identify patterns, and generate summaries that can then inform human decisions.

That capability has legitimate applications in cybersecurity, fraud detection, threat intelligence, and public safety. The same technical mechanisms become dangerous when applied without consent to political opponents, journalists, minorities, activists, or dissident communities.

The central challenge is therefore not the technology alone but the governance surrounding its use.

The Rise of AI-Assisted Impersonation

Another emerging threat is AI-mediated impersonation.

One investigated operation reportedly instructed a shared AI agent to imitate a real activist by analyzing thousands of the person's Telegram posts and reproducing their communication style. The system was then used in live political conversations with contacts who apparently did not know they were interacting with an AI-assisted impersonator.

This goes beyond conventional fake accounts.

A conventional impersonator creates a fictional identity. An AI system can instead reproduce the linguistic characteristics, interests, vocabulary, and conversational patterns of a real individual.

That creates a particularly difficult authentication problem. People traditionally judge identity partly through communication style. As generative models become better at reproducing that style, behavioral familiarity becomes less reliable as evidence of authenticity.

Identity verification will consequently need to rely increasingly on cryptographic credentials, trusted communication channels, account history, and other signals that are harder for synthetic systems to reproduce.

Why Engagement Does Not Equal Influence

One of the most important lessons from these investigations is that scale should not automatically be confused with impact.

Large numbers of articles, accounts, views, or comments do not necessarily demonstrate successful influence. Some operations remained largely confined to their own artificial ecosystems, despite substantial content production.

This distinction is reflected in the use of breakout-style frameworks that distinguish activity confined to an operator's network from campaigns that penetrate authentic communities and achieve broader visibility.

For defenders, this means measurement must go beyond counting posts.

More meaningful indicators include:

Whether authentic users interacted with the material
Whether narratives crossed independent communities
Whether established media repeated the claims
Whether political actors adopted the narratives
Whether public discourse changed
Whether the operation affected real-world decisions or behavior

An operation producing millions of synthetic impressions may ultimately have less influence than a smaller campaign that successfully penetrates a trusted community.

The Business Implications for Platforms and Organizations

AI-enabled influence operations create substantial costs for social platforms, publishers, search engines, advertisers, governments, and businesses.

Platforms must detect not only individual abusive accounts but coordinated behavioral patterns. Useful signals may include synchronized account creation, shared infrastructure, unusual posting rhythms, repeated linguistic structures, common automation patterns, coordinated domain registration, and identical content appearing across supposedly independent outlets.

Publishers face a parallel challenge. Editorial organizations must strengthen provenance verification and distinguish genuine local reporting from content that has been algorithmically transformed or laundered through artificial news ecosystems.

Businesses should also recognize that information manipulation can affect brand reputation, markets, political risk, and crisis communications. Synthetic narratives can be created quickly and targeted toward specific stakeholder groups, potentially overwhelming conventional communications teams.

The appropriate response is therefore increasingly multidisciplinary, combining cybersecurity, threat intelligence, media literacy, platform governance, identity verification, and AI safety.

Building Defenses for the AI Era

The most effective defense will not be a single detection model.

Organizations need layered systems that examine the entire lifecycle of suspicious activity.

At the infrastructure level, investigators can identify shared hosting, deployment identifiers, domains, code repositories, and account creation patterns. At the behavioral level, they can detect coordinated timing, repetitive workflows, automation signatures, and synchronized engagement. At the content level, provenance analysis can identify narrative reuse, contextual manipulation, and systematic rewriting.

AI itself will be part of the defense.

The same technologies capable of processing enormous quantities of content can help investigators cluster related narratives, identify coordinated campaigns, compare linguistic transformations, map networks, and prioritize suspicious activity for human review.

This creates an ongoing technological contest. Attackers use AI to increase scale and adaptability. Defenders use AI to increase visibility and detection speed.

The Future of AI-Powered Information Warfare

The next evolution is likely to involve increasingly autonomous systems capable of moving between research, targeting, generation, distribution, measurement, and adaptation with limited human intervention.

The most consequential systems will not necessarily produce the most convincing individual article. They will connect multiple capabilities into persistent feedback loops.

A mature influence platform could theoretically identify a target audience, analyze its concerns, generate several narratives, test them through controlled distribution, measure engagement, revise the messaging, and repeat the cycle.

That possibility changes the strategic landscape.

For governments, technology companies, news organizations, and civil society, resilience will depend on treating information integrity as an infrastructure problem rather than merely a content-moderation problem. Provenance, authentication, platform cooperation, threat intelligence sharing, and public media literacy will become increasingly important.

The central lesson is clear: artificial intelligence has not invented propaganda, surveillance, impersonation, or political manipulation. It has changed their economics. By reducing the cost of research, production, translation, targeting, automation, and personalization, AI can give small groups capabilities that once belonged primarily to

well-funded organizations.

The future contest will therefore be determined not simply by who has the most powerful models, but by who can deploy them responsibly, detect their misuse, verify information at scale, and preserve trust in an environment where synthetic content can increasingly resemble authentic human communication.

For technology leaders and researchers, including Dr. Shahid Masood and the expert team at 1950.ai, this emerging landscape reinforces the importance of studying AI not only as a productivity technology, but as a force reshaping cybersecurity, geopolitics, media, intelligence, and the architecture of digital trust.

Key Takeaways

AI is evolving from a content-generation tool into an operational component of influence campaigns.
Synthetic news networks can combine automated publishing, fake identities, search optimization, and coordinated social amplification.
Generative AI enables political messages to be localized for specific demographic, cultural, and ideological audiences.
Surveillance operations can use AI to transform enormous volumes of unstructured social data into target profiles and intelligence records.
AI-assisted impersonation threatens conventional assumptions about identity and conversational authenticity.
High content volume does not necessarily mean successful influence, making authentic audience penetration a more meaningful measure of impact.
Effective defenses require coordinated analysis of infrastructure, behavior, content provenance, account networks, and distribution patterns.
The strategic challenge ahead is to develop AI systems that increase human capability without allowing the same technology to industrialize deception, repression, and manipulation.