Pangram and Redis enhance visual detection and memory for AI agents
Pangram adds video to its AI-generated image detection. Redis extends Iris with routing, memory, and audit modules for agents.
Pangram presented on November 4, 2026, two vision models to detect AI-generated content, while Redis expanded Iris with six capabilities aimed at agents. In both cases, the challenge is the same: better control over what AI sees, decides, and retains, without relying on a single model for the entire chain. Pangram also adds native video uploads, a Chrome extension, and an API; Redis focuses on routing, memory, and observability. All this arrives as deployments move out of the lab and practical limits quickly become visible.
Detect an image, then track the video frame by frame
Pangram Image 1.0, the main model, detects AI-generated images and displays a heatmap, that is, a map of suspicious areas within an image. Its smaller sibling, Pangram Image Flash 1.0, tackles videos by analyzing each frame separately. According to the company, Image 1.0 identifies 99.2% of AI-generated images with a false positive rate of 0.27%.
The promise is useful, but the measurement matters as much as the announcement. Pangram says it tested Image 1.0 on a set of 1,200 synthetic images built from 100 real photos and 12 image generators, with an accuracy of 99.96% on this benchmark designed by the company. This performance does not represent a universal rate; it describes a targeted test on close real/synthetic pairs.
For video, Pangram followed a similar logic: 129 real videos from social networks, 14 video generators, and 1,529 videos in total. The DF26 benchmark, cited by the company for Flash 1.0, includes 271 real videos and 2,420 synthetic videos produced by seven recent models. The authors of this benchmark say that humans as well as deepfake detectors perform near chance on this terrain [to be verified], giving an idea of the difficulty level.
Quand la détection sort du labo, les détails font la différence
Pangram claims that Flash 1.0 reaches 98.7% on DF26, 29 points higher than the next detector. The model also appears in the dashboard via native video uploads, allowing detection of which part of a video contains AI detection. The Chrome extension further expands usage to social media sites, shopping, or rental platforms. Convenient, then. And a bit less comfortable for those who relied on ambiguity.
Version 1.0 also improves support for deepfakes, low-resolution images, and the latest generation models, according to Pangram. The minimum accepted size drops to 256 x 256 pixels, but images under 512 pixels trigger a reduced confidence warning. The company also says it trained the model to recognize transformations introduced by compression, resizing, blurring, or noise, whether from platforms or evasion attempts.
This evolution matters mainly because it changes the entry point. In July 2026, the research preview accepted only images of at least 512 x 512 pixels and was not considered reliable on deepfakes or face swaps. Now, Pangram moves from a research prototype to a more widely distributable tool, with image and video API included.
Redis veut éviter aux agents de bricoler leur mémoire à la main
Redis, for its part, has extended Iris, its real-time context engine, with six composable capabilities for AI agents. Four are generally available, two remain in beta. The idea is simple: not to force developers to assemble several separate bricks to give an agent memory, context, routing, live data, and communication mechanisms.
Intent Router first interprets the meaning of a request rather than its exact wording. It can send it to a memorized FAQ, an agent skill, a human, or an LLM, with conditional rules and safeguards. Agent Playbooks serve to replay established procedures when the answer already exists, while reserving new executions for cases requiring fresh data. Redis says this logic reduces resolution time and token consumption as playbooks are reused.
Agent Memory relies on MemoryETL to build memories from user actions, documents, and semi-structured knowledge bases. Memory is organized hierarchically and accompanied by inspection, audit, and provenance functions. Most importantly, it can also delete data it should not retain, including personally identifiable information. Context Gateway provides a governed entry point to connected context sources, supporting MCP servers and tool caching. Redis also sees this as a way to limit token costs.
Le vrai sujet : tenir en production sans perdre le fil
Both announcements tell the same story from different angles. Pangram works on visual proof, Redis on operational continuity. In both cases, the problem is no longer to demonstrate that a technology exists, but to know if it holds up when contents look alike, sessions are interrupted, and data piles up. That’s when products become serious.
Redis also adds Lens, to trace the origin of a decision, and Agent Mesh, to manage connection interruptions between agents. Lens links a result to the skill, router, memory, and original session, with success, latency, and token metrics. Agent Mesh covers ephemeral streams, automatic recovery, agent handoffs, and durable SSE flows. These features are still in beta, unlike the memory, routing, and context bricks already generally available.
Redis recalls that Iris started in May 2026 as a context and memory layer between an agent and the data it depends on. The new set expands this role to full coordination of a request, from interpretation to reuse. Pangram, for its part, pushes detection beyond static images, with a more concrete promise for moderation, commerce, and visual verification uses. For product teams, the question is no longer just “does it work?”, but “where does it break, and how fast?”.
Key points
- Pangram launched Image 1.0 and Image Flash 1.0 on November 4, 2026.
- Redis Iris adds six capabilities, four of which are generally available.
- DF26 includes 271 real videos and 2,420 synthetic ones.
- Pangram announces 98.7% on DF26 for Flash 1.0.
- Redis offers 5,000 free messages at startup.
- Image 1.0 accepts images from 256 x 256 pixels.
By the numbers
- 99.2% — rate announced by Pangram for AI image detection, November 2026.
- 0.27% — false positives announced for Image 1.0, November 2026.
- 98.7% — score claimed by Flash 1.0 on DF26.
- 5,000 messages — startup offer announced by Redis for Iris.
- 256 x 256 pixels — minimum entry threshold for Image 1.0.