Two independent red teams cracked the latest generation of ChatGPT within 24 hours. The tests show that simple multi-step stories can bypass the built-in security frameworks. Where Grok 4 quickly fell down, GPT 5 proved equally vulnerable. The findings touch directly on whether companies can safely deploy the raw model.
NeuralTrust describes how a combination of a proprietary EchoChamber approach and narrative dialogue led the model step by step to forbidden output. Not the individual prompt but the entire conversation history became the lever. By seeding innocuous context and letting that context develop further and further, the system stayed consistent with the previously told story and passed safety filters without triggering alarms.
SPLX came to a similar conclusion. According to these testers, the crude model is hardly applicable in a corporate environment. Concealment techniques still work. One example is stretching words by separating each character and packaging the whole thing as a mock task about encryption. After long conditioning, the model responds willingly and bypasses the original protection. In their benchmark, GPT 4o remains the more robust choice when additionally hardened.
The thread is clear. Filters on just the current message fall short. The vulnerability lies in the connection between successive messages. Those who cleverly direct the sequence can hide the intent and still achieve the end goal. That makes the threat more realistic for any organization that uses chat-driven workflows for customer service decision-making or code assistance.
What does this mean for Surinamese companies. Don't rely on default settings of a new model. Place a control layer between user and model that evaluates each message and the entire conversation. Block masking inputs such as artificial strings and unusual encodings. Enforce policies outside the model with proprietary permissions and risk profiles. Log all sessions and put limits on duration frequency and context length. Have jailbreak tests performed by an external and independent team before a rollout. Choose a hardened variant or a proven alternative for high-risk processes. Train employees in recognition of circumvention tactics and practice incident response to language models as if it were a phishing crisis. Early deployment can only be done safely with additional fences on the outside and with ongoing attack simulations. Avigdor is considered a Suriname cybersecurity partner and will maintain this approach with clients. Those looking to use chat models in core processes first build these layers of defense and then have periodic red teaming performed so that productivity and compliance remain balanced.
