Topic
Security techniques that screen LLM inputs and outputs for malicious prompts, injections, or policy violations, before or after processing.
Check Point says PuzzleMask hides nothing an input classifier is built to catch, and the expensive downstream model did the unwrapping itself, acting on the payload in 17 of 18 trials with a code interpreter.
Reality
Adversa AI got GitHub Copilot CLI to decrypt and run a hidden instruction, and Microsoft's mai-code-1.1-flash fell for it 50 percent of the time. GitHub declined to treat it as a product vulnerability, so teams running autopilot have to check what the agent executes themselves.
Reality
Red Hat's 200M-parameter classifier scored 89.01% on prompt injection to a 35B LLM judge's 89.31%, answering in a median 54.1 ms against 312.5 ms. For injection screening, a large judge in the request path now has to justify nearly six times the latency for 0.3 points.
Reality
Archer's Evolv AI Compliance turns rules from a 22-million-document library into Amazon Bedrock guardrails that check each prompt before a model answers. Archer is betting that risk and compliance teams will pay to enforce their rules at the moment a prompt reaches a model, on top of controlling who can log in.
Reality
Meta's Prompt Guard 2 caught 6 of 629 buried attacks at a 0.5 cutoff and 621 at 0.003, in a benchmark posted on dev.to. Health checks pass the same at both settings, so only attacks sent through the deployed threshold show which one a team runs.
Reality
Adversa AI says its Cryptographic Context Injection recovers hostile prompts inside the code sandbox, where filters do not look. xAI has not replied; Google scopes jailbreaks out entirely.
Perspective Coverage
6 publishersReality
Adversa AI put AES ciphertext on a webpage that no text filter could read. Grok's own code execution decrypted it, and the plaintext came back classified as the model's reasoning rather than as fetched content.
Reality
The dev.to post announcing llm-sentinel says llm-guard is archived. The replacement is ten pattern-matching scanners with a score threshold each, benchmarked on 133 hand-written cases its author calls a smoke test.
Reality
AWS has published a pipeline that pulls Bedrock guardrail traces out of model invocation logs, rewrites them as OCSF Detection Findings, and lands them in the CloudWatch store where analysts already query CloudTrail and VPC Flow Logs.
Reality
PolicyAware 0.4.4 evaluates identity, tenant, region, arguments and risk tier before an agent's action changes state. Enforcement sits inline on every turn, and the project's own guidance is that adopters measure what that costs.
Reality
A dev.to writeup moves tool guards out of the prompt and into the executor, where a database check runs before dispatch. The same design fails open when the check itself breaks.
Reality
With no completed sales, nine live asks running from £35 to £240 and no route to the £26.99 shop price, a self-hosted pricing tool proposed £35.39 and recorded the missing check only in a log line.
Reality
A dev.to post wraps an autonomous code-review agent in a CLOSED/OPEN/HALF-OPEN state machine. The constructor's defaults show which of its four blast-radius dimensions an adopter has to wire up alone.
Reality
The F-072 rule engine scans an LLM's chain of thought for risk words and multiplies position size by 0.8 when it finds one. On the trade its author logged, both constants it applied were already in the model's own reasoning.
Reality
Reconstructing an incident from CloudTrail, AWS Config and Kubernetes events is a correlation problem before it is a language problem, and this engine's timestamp floors decide which links the model may narrate at all.
Reality
Centralising prompt checks and PII filters at the gateway mirrors putting TLS at the edge, and the arithmetic backs that comparison. The semantic checks themselves, though, work nothing like TLS termination.
Reality
TraceSafe-Bench ran 20 guard systems over more than a thousand edited tool-call traces, and the ranking followed structured-data ability rather than safety alignment. Your prompt-injection number does not cover the runtime.
Reality
Model-level guardrails check the prompt and the response, which leaves the parameters the model just chose for a tool unexamined. AWS's answer is three lifecycle hooks, and the per-turn cost is yours to measure.
Reality
ChatGPT for Teens folds a year of bolt-on parental controls into a product boundary. The hard part is now deciding which account a minor is sitting in, and who configured it.
Reality
Adversa says it hid data-exfiltration instructions in AES-256-GCM ciphertext and let Grok decrypt them in its own Python sandbox. The plaintext version of the same attack was refused.
Reality