Subregions in the ventromedial prefrontal cortex integrate threat and protective information to meta-represent safety

In their natural habitat, the Sundarbans Tiger, which is a formidable threat to humans, is justifiably feared. However, with a gun in hand, we fear the Tiger less. At the zoo, behind the protection of laminated glass, fear is replaced with enthused curiosity. In all of these scenarios, the sensory features of the Tiger remain stable, yet the perception of safety fluctuates. These fluctuations in safety occur as a function of information unrelated to the Tiger’s features, but changes in the perception of safety. Despite this knowledge, the contributions of safety estimation beyond external threats are largely ignored in the existing literature. To accurately estimate safety, we need to integrate threat-related information with information about our ability to protect ourselves [1,2]. From an evolutionary perspective, these factors determine how likely we are to succeed in surviving encounters with natural dangers.

The ability to recognize safety is critical for adjusting adaptive defensive responses, reducing stress, and initiating other survival behaviors, including foraging and mating [3,4]. How the brain integrates multiple sources of dynamic information to compute safety estimates remains to be identified. One theory is that the coding of safety is constructed based on representations that reflect the learned and phenotypic features of external threats [5–7]. The brain creates an internal model that integrates incoming sensory information about a stimulus with other relevant information such as context (i.e., Am I at the zoo or alone in the forest?). This “other” information does not pertain to the threat itself but is important for accurately estimating safety. As relevant information changes so can the safety estimate, even if the threat itself remains unchanged. Similarly, changing threat features (i.e., Is the tiger awake or asleep?) should trigger updated safety estimates through integration in safety circuits with consequences for survival behavior. Such Meta-representations are likely to involve neuronal ensembles that integrate the valuation of threats with information about our ability to protect against them [8,9]. Depending on the resulting calculation, defensive excitation or inhibition occurs [10–13].

We test 2 primary hypotheses regarding the functioning of safety neural circuitry. First, we test the hypothesis that the neural systems involved in representing threat and protection are dissociable during Safety Prediction [2]. We extend beyond models focused on the external environment to examine fluctuations in safety as they relate to self-relevant states (e.g., the value of protection). We argue that self-relevant states—the extent to which one can successfully protect oneself—are a crucial, but overlooked, factor in estimating safety. Going face-to-face with a tiger with no weapon in hand will likely result in death. Yet, with a knife, we have an increased chance of deterring the predator and surviving. A gun further increases survival likelihood and turns the tables on the tiger as the likely casualty. Second, we hypothesize that the brain integrates threat and protective information to confer a safety “meta-representation” [14–16]. To test this, we manipulated the safety value of threat and protective stimuli as a function of their pairing without altering the perceptual features of either stimulus. For example, a tiger becomes less dangerous when faced with a gun as opposed to a knife, but the tiger itself does not change. We hypothesize that the safety modulator (e.g., the gun or knife) will be meta-represented during the evaluation of the modulated stimulus (e.g., the tiger), even though the modulator is not being displayed and therefore not visually perceived.

We hypothesize a candidate region for human safety coding is the ventromedial prefrontal cortex (vmPFC). Recent theoretical developments suggest representations of threat and protective information are encoded in canonical defensive and cognitive neural circuits, the latter including the vmPFC [2,17–22]. The vmPFC aids the affective processing of safety signals as well as the acquisition of new threat associations and threat extinction during Pavlovian threat conditioning [2,17–28]. With relevance to the current study, we propose that external threats are computed in a bottom-up fashion, primarily driven by sensory processing regions of the brain, whereas protection is integrated with threat information through top-down metacognitive circuitry related to self-evaluation, including the vmPFC [29–31]. Work outside of the threat context points to the vmPFC as pivotal in supporting decisions related to the self [16] and in showing bias toward information concerning the self [31]. The vmPFC contributes to binding self-relevant information across cognitive domains such as perception, memory, and decision-making, thereby enhancing the integration of stimuli with self-representations [32]. Research on threat controllability provides links between self-referential cognition and safety—controllability, which relies on self-referential processing, improves fear extinction, which reflects safety, in humans, thus implying a role of the self in safety estimation [17–28,33]. Consequently, we identify the vmPFC is a potentially important region for integrating self and external information to formulate safety estimations.

Although we propose that successful safety estimation in humans relies on integrating multiple distinct components, how the brain interprets and weighs these factors remains unknown. No prior study to our knowledge has systematically manipulated both threat and protection concerning safety judgments. In 2 preregistered samples using a novel Safety Estimation Task (Fig 1A) (N = 100 behavioral; N = 30 functional magnetic resonance imaging, fMRI; https://osf.io/hw3r9), we examined how subjects evaluated different types of safety information (Safety Prediction), as well as how they integrated information to estimate safety when perceptually identical stimuli changed in value (Safety Meta-representation). As a comparison, we examined neural activation when safety was certain during the outcome phase (Safety Recognition). We also tested how safety estimation changed as a function of experience (Safety Value Updating) in a separate task administered before and after the Safety Estimation Task. We examined safety coding at the whole-brain level but focused on the contributions of the vmPFC as a hypothesized safety coding hub.

Fig 1. Safety Estimation Task and Safety Value Updating Task schematics, including safety probabilities, a conceptual model of safety estimation and integration, and MRI session procedure details.

(A) Safety Estimation Task. An example trial is presented. Subjects were told to imagine they were battling dangerous animals with powerful weapons. On each trial, subjects saw stimuli pairs comprised of a threat (animal) and protection (weapon) with presentation of threat/protection counterbalanced. First a weapon or animal was presented (Safety Prediction) and subjects made an initial estimation of whether they would win or lose the battle, responding with a button press. Then, the paired stimulus was presented (Safety Meta-representation) and subjects made an updated judgment as to whether they would win or lose, responding with a button press. After both stimuli were presented, subjects saw the outcome of the battle depicted as either a shock (loss) or no shock (win) (Safety Recognition). If subjects lost and received the shock image, they had a 20% chance of receiving an electric shock to the wrist. (B) Safety probabilities. Threat and protection stimuli were each set on a four-point continuum with equivalent experimentally established safety probabilities. Italics depict the average shock value for each stimulus across all pairings. Paired probabilities are depicted in the heatmap. Probabilities were experimentally established prior to testing and were not made known to participants although the continua were easily identified based on prior knowledge and outcomes during the task. (C) MRI session procedure. On the first day, subjects completed the naive version of the Safety Value Updating Task. At this point, subjects had not completed instructions for the Safety Estimation Task and thus had no knowledge of the task or stimuli relevance. Subjects then completed a structural scan, during which they learned about the Safety Estimation Task. After completing 5 practice trials, subjects completed Run 1 and Run 2 of the Safety Estimation Task. An average of 26 min elapsed from the start of Run 1 to the start of Run 2. On the second day, subjects completed Run 3 and Run 4 of the Safety Estimation Task, with an average of 25 min from the start of Run 3 to the start of Run 4. After Run 4 ended, subjects completed the knowledgeable version of the Safety Value Updating Task. Day 2 of testing took place on average 1 day and 7 h after day 1. Task images are approximate reproductions. Source data can be found at https://osf.io/8qg7y/ under “Behavioral data”.

https://doi.org/10.1371/journal.pbio.3002986.g001

During the Safety Estimation Task, subjects were shown stimuli pairs comprised of an external threat (dangerous animal) and a self-relevant protection (powerful weapon). Four stimuli for each threat and protection were used (Fig 1B). Presentation of stimuli was counterbalanced such that in some trials, the protection was shown first and in other trials, the protection was shown second. Subjects made binary forced-choice judgments about whether they thought they would win or lose the battle against the dangerous animal using the weapon they were provided. All trials included a threat and protection, separately presented to allow for analyzing subject response when shown the first stimulus (Safety Prediction) separate from the second stimulus (Safety Meta-representation). Safety probabilities for threat and protection stimuli were matched such that high safe protection and high safe threat were equally likely to result in a win. Combinations of threat and protection stimuli were also matched with safety outcomes varying as a function of the average safety value of each stimulus in the pair. After both stimuli were presented, subjects saw the outcome of the battle (Safety Recognition). Lost battles risked delivery of an electric shock to the subject’s wrist (randomly delivered on 20% of lost trials). Successful battles resulted in 100% safety with no electric shock.

Results are reported as facets of safety estimation, (1) Safety Prediction in response to the first stimulus presentation during the Safety Estimation Task when subjects could estimate safety based only on partial information; (2) Safety Meta-representation at the second stimulus presentation during the Safety Estimation Task when subjects could estimate safety on full information (Fig 1C); (3) Safety Recognition at the outcome during the Safety Estimation Task when subjects knew whether they were at risk of electric shock; and (4) Safety Value Updating comparing the “knowledgeable” and “naive” rounds of the passive viewing task. For each section, behavioral models are reported as well as univariate and multivariate fMRI results. Safety value (high versus low) and safety relevance (external versus self) are both considered.

Safety prediction

Differences in Safety Prediction represented a bias toward initial safety information as a function of relevance (external versus self), given that safety probabilities for threat and protection were identical.

Behaviorally, subjects in both the behavioral and MRI samples tracked probabilities of winning and losing across the safety continuum such that subjects estimated a higher probability of winning for stimuli with higher safety probabilities (Fig 2A; Table A in S1 Text). Safety relevance (external versus self) affected initial safety bias such that subjects estimated greater safety when the first stimulus presented was protection, behavioral sample difference between protection and threat α = 0.23, 95% CI [0.21, 0.25]; MRI sample α = 0.28, 95% CI [0.25, 0.31] (Fig 2B).

Fig 2. Behavioral results showing safety estimation and biased predictions based on stimulus type.

(A) Mixed effects logistic regression depicting the association between experimentally established safety (x-axis) and subjective safety estimate derived from subject ratings for win/lose during battles (y-axis; 0 = lose, 1 = win). Threat and protection safety continuums are equivalent (x-axis). Results indicated subjects differentiated safety in accordance with the experimentally established safety continuum and tracked safety probabilities across the safety continuum for both stimulus types. See Table A in S1 Text. (B) Psychometric curves were fit for Safety Prediction in both samples. Subjects reached the safety detection threshold (α) faster when protection was presented as the first stimulus in the battle pair. Behavioral N = 100 (left), MRI N = 30 (right). (C) Subjects were more likely to modify their safety estimations during Safety Meta-representation when threat stimuli were presented first followed by protective stimuli, especially when the safety value of the second stimulus changed to increase safety probability. Source data can be found at https://osf.io/8qg7y/ under “Behavioral data”.

https://doi.org/10.1371/journal.pbio.3002986.g002

Univariate fMRI analyses showed that the brain tracked safety value in response to the first stimulus presented, with dissociable circuits activated depending on safety value. As the stimuli increased in safety value, so did activation in the vmPFC, as hypothesized (Fig 4A and Table B in S1 Text). Considering each stimulus type separately, increasing safety for threatening animals activated the lateral occipital cortex (Fig 4B), and increasing safety for protection activated the vmPFC, amygdala, temporal pole, and anterior cingulate cortex (ACC) (Fig 4C).

Activation corresponding to decreases in safety was observed in the occipital pole and postcentral gyrus (all stimuli and threatening animals, Fig A panels A and B in S1 Text). Decreasing the safety value of protection did not show significant differences in parametric activation (Fig A panel C in S1 Text).

Informational Connectivity testing multi-voxel pattern synchronization between regions of interest (ROIs, Fig 6A) revealed a dense network involved in coding Safety Prediction (Fig 6B). The ACC (caudodorsal), insula (anterior), striatum (dorsal and ventral), vmPFC (anterior and posterior), thalamus, and hippocampus were connected during safety decoding. Betweenness Centrality revealed the ACC as the primary hub of the network, with the insula, dorsal striatum, and posterior vmPFC also showing centrality contributions. While encoding the safety value of protective weapons, the anterior vmPFC emerged as the central hub, and the thalamus was eliminated as part of the network. While encoding safety in response to threatening animals, the posterior vmPFC emerged as the network hub and connections with the hippocampus, thalamus, and ventral striatum were no longer significant.

Safety meta-representation

During Safety Meta-representation, stimuli were examined in terms of safety fluctuation: when the same stimulus had a higher-than-average safety value (e.g., a fist paired with a cat) compared with a lower-than-average safety value (e.g., a fist paired with a grizzly). Fluctuation in neural response reflects the integration of information about the initial stimulus encountered (e.g., cat, grizzly) with the second stimulus presented (e.g., fist) while holding the perceptual experience of the second stimulus constant.

Subjects’ safety meta-representation behavior was consistent with their predictive behavior such that subjects in both samples estimated a higher probability of winning for stimuli with higher safety probabilities (Fig 2A and Table A in S1 Text). There was no significant difference in the threshold of safety detection as a function of the second stimulus type, behavioral sample difference between protection and threat α = 0.01, 95% CI [−0.03, 0.04]; MRI sample α = −0.02, 95% CI [−0.09, 0.04].

During Safety Meta-representation, subjects fixated more on the initial safety value if protection stimuli were presented first. In other words, subjects were more likely to switch their safety prediction from win to lose or lose to win when threats were presented first and protection second (stimulus type: behavioral B = 0.23, SE = 0.02, z = 11.20, p SE = 0.03, z = 4.04, p SE = 0.006, z = 6.84, p SE = 0.01, z = 9.59, p Fig 2C). Overall, subjects rated winning probabilities as higher if the second stimulus was a powerful weapon following a dangerous animal, compared to a weak weapon following a safe animal (75% versus 53%), despite equivalent experimentally established safety probabilities (Fig 3A). However, subjects rated safety probabilities as equivalent when a dangerous animal followed a weak weapon or a safe animal followed a powerful weapon (63% for both) (Fig 3B).

When stimuli increased in safety value compared to its average, vmPFC activation increased parametrically (Fig 4D and Table B in S1 Text), evincing a more distributed pattern of activation across the vmPFC compared with Safety Prediction. Considering each stimulus type separately, neural response to external threats increased in safety (as a function of being paired with more powerful weapons) activation increased in the vmPFC and lateral occipital cortex (Fig 4E). As self-relevant protective stimuli increased safety, neural response increased in the occipital cortex (Fig 4F). Conjunction analyses identified the vmPFC as a common neural substrate of Safety Prediction and Meta-representation (Fig 4G).

Fig 3. Behavioral results showing safety estimation updating as a function of stimulus presentation order.

(A) When threat was presented first and protection second, subjects were overconfident about their safety estimates when armed with a high-value weapon. (B) When protection was presented first and threat second, subjects were more likely to differentiate according to the threat pairing in line with experimental probabilities. Heatmap scales for subjective safety ratings are normalized for responses in each of the Behavioral and MRI samples. Source data can be found at https://osf.io/8qg7y/ under “Behavioral data”.

https://doi.org/10.1371/journal.pbio.3002986.g003

Fig 4. Neural response to safety increases during each task phase, highlighting the role of the vmPFC in responding to safety increases and protection stimuli.

Analyses in were conducted using FSL Randomise, TFCE, FWE-corrected p A–C) Parametric increases in whole-brain neural activity that track increases in experimentally established safety value of stimuli during Safety Prediction. The first stimulus presented represented a bias to partial information, which measures a differentiation in neural activity as a function of stimulus type (threat versus protection). Significant clusters indicate activation increased in those regions as safety probability increased. Safety increase was based on the average experimentally established safety probability of each stimulus (protection continuum order: fist, stick, gun grenade; threat continuum order: cat, goose, lion, grizzly). (A) Threat and Protection collapsed, (B) Threat only, (C) Protection only. (D–F) Parametric increases in whole-brain neural activity that track increases in experimentally established safety value of stimuli during Safety Meta-representation. The second stimulus safety value was based on the combined safety probability of the first and second stimuli. For analyses, safety was based on comparison with the average safety value of the stimulus. For example, if a stick was shown as the second stimulus and was paired with a cat, the probability of safety would increase from 35.72% (safety average for all stick trials) to 57.14% (safety when stick is paired with cat) (see Fig 1B). (D) Threat and Protection collapsed, (E) Threat only, (F) Protection only. (G) Conjunction of Shared Activation between Safety Prediction and Safety Meta-representation. Conjunction analyses for the safety prediction phase of increasing safety versus increasing danger and shared activation with the safety meta-representation phase of increasing safety versus increasing danger. All stimuli represented. Results indicate overlapping activation in the vmPFC; Z = 2.3, p Fig 1D). Analyses examined the contrast of the Knowledgeable version (post-task viewing) versus the Naive version (pre-task viewing) (E). Searchlight was a priori restricted to the vmPFC using an ROI defined via Neurovault. Source data can be found at https://osf.io/8qg7y/ under “MRI data.” ROI, region of interest; vmPFC, ventromedial prefrontal cortex.

https://doi.org/10.1371/journal.pbio.3002986.g004

In response to increasing danger, activation increased in the insula, thalamus, ACC, and PAG (Fig A panel D in S1 Text), showing a more typical pattern of defensive circuitry activation than that observed in response to dangerous stimuli at first presentation. As threatening animals become less safe, the bilateral thalamus, right insula, pre-supplementary motor cortex, and medial occipital cortex parametrically responded with increased activation (Fig A panel E in S1 Text). When self-protective stimuli were rendered less safe than average because of the initial danger value of the threat, neural response increased in the insula, thalamus, and ACC (Fig A panel F in S1 Text).

Testing multi-voxel pattern synchronization using Informational Connectivity, the same safety network was decoded with the addition of the amygdala in response to Safety Meta-representation (Fig 6C). Informational Connectivity during Safety Meta-representation tracked changes in safety from overall average values, in line with univariate results. Betweenness Centrality revealed a switch from the ACC as the top Safety Prediction hub to the dorsal striatum as the top Safety Meta-representation hub when examining all stimuli and in response to threatening animals, with the ACC, insula, and posterior vmPFC playing centrality roles. The posterior vmPFC was the central hub for the Meta-representation of protective weapons, followed by the dorsal striatum and ACC and the insula no longer showing as a central hub.

During Safety Prediction, subjects were quicker to detect safety when presented with self-relevant protective stimuli compared to when presented with externally relevant threat stimuli. Neurally, vmPFC activation parametrically increased as protection increased in safety value. Threat stimuli, in contrast, activated sensory and defensive neural systems. Despite equivalent experimentally established safety probabilities for threat and protection stimuli, only protection evoked activation in the vmPFC. One intriguing possibility for vmPFC sensitivity to protective weapons, beyond their safety relevance to the self, is that the vmPFC was coding ownership associations between objects and the self [31]. We did not directly test whether ownership was part of the cognitive milieu in our study, but our findings are consistent with the perspective that the vmPFC assigns personal significance to objects based on its meaning and function for the self [35]. Our findings agree with prior literature establishing that the vmPFC tracks self-related associations beyond reward-based learning [35] and point to a specialized role of the vmPFC in tracking self-relevant information during safety estimation.

As threat stimuli increased in danger during Safety Prediction, so did activation in the occipital pole and postcentral gyrus. We interpret this activation to reflect several processes related to visual, attention, and sensory processing when evaluating stimuli with heightened salience, like those representing a threat. The occipital pole is involved when visual processing demands are enhanced, as is the case when evaluating threatening stimuli that capture attention more effectively—as is shown in prior work suggesting threatening stimuli are prioritized in the visual system [36,37]. The postcentral gyrus contains the primary somatosensory cortex and is involved in processing somatic sensory information, which may reflect an embodied response to perceived threat preparing for potential defensive action [38]. Enhanced sensory processing in both visual and somatosensory areas may thus serve as an adaptive mechanism to improve threat detection and prepare for action.

During Safety Meta-representation, subjects were again quicker to detect safety value for protection. Neurally, subjects meta-represented the first stimulus when evaluating the second stimulus, despite the absence of perceptual information about the first stimulus [34,39]. In response to threat stimuli, the same vmPFC region that activated to protection during Safety Prediction was activated. In response to protection stimuli, the same sensory regions of the visual cortex that activated to threat during Safety Prediction were activated. In other words, the pattern of activation at the second stimulus was inverted, which we interpret as meta-representation during safety integration. This process of reactivation/replay has been observed in aversive learning in humans [34,40]. We observed patterns of activation at a whole-brain level, demonstrating striking similarities to the initial state encountered. We did not conceptualize meta-representation as replay in our study because our task does not afford the typical planning opportunity that is tested during traditional examinations of replay. Our interpretation is bolstered by the task design: stimuli at the second stimulus pairing were perceptually identical and only differed in safety value as a function of their pair. Thus, the neural systems responding to increases in safety during Meta-representation were responding to changes in the safety value as a function of the first stimulus safety value and its resulting influence on the overall safety probability. Similar behavioral patterns emerged: Subjects were more likely to update predictions during Safety Meta-representation when self-relevant protection conflicted with initial external threat information. On the other hand, when protection was presented first, subjects did not shift safety estimations. Our results fit with a broader body of literature showing that memory retrieval induces aspects of the pattern of neural activity evoked by the original stimulus presentation [39,41].

Our findings converge with prior work indicating the vmPFC is involved in updating mental representations of threat and reward particularly in contexts like fear extinction [17]. Given vmPFC involvement in the physiological regulation of autonomic processes like heart rate and skin conductance [42,43], it is likely that vmPFC engagement we saw during our task reflects a more complicated myriad of processes that we did not explicitly measure, including emotion regulation [44,45]. Additionally, our behavioral results show safety recognition varied as a function of stimulus type, and the vmPFC was preferentially activated to protection during the Safety Estimation Task. These findings are consistent with the vmPFC playing a role in self-relevant processes specifically rather than undifferentiated emotion regulation broadly. If the vmPFC were primarily tracking the latter, we would expect to see consistent activation regardless of stimulus type because both stimuli presentations occur during a need for regulatory engagement. Our findings fit with an interpretation of the vmPFC as providing a substrate for linking experiences, contexts, and events to the bio-regulatory or emotional state of the self [32,46].

Given the lack of information that classic univariate connectivity approaches have (i.e., PPI), we used informational connectivity to examine synchronization of voxels. Multivariate connectivity revealed a safety network consisting of the anterior and posterior vmPFC, dorsal and ventral striatum, caudodorsal ACC, and insula. The hippocampus, thalamus, and amygdala also emerged as connected to this core network, but with less consistency depending on stimulus relevance. In response to Safety Prediction, threat and protection networks showed a shift in hub organization from the posterior to anterior vmPFC. The dorsal striatum emerged as a core hub for Safety Meta-representation. The dorsal striatum has been previously linked to punishment-based avoidance, with dorsal striatum damage resulting in suboptimal defensive choice [47]. We situate our findings with consideration of this prior work and interpret Safety Meta-representation as necessary to generate choices about defensive action. The ventral striatum and hippocampus were both connected to the safety network during protection evaluation but not during threat evaluation. The vmPFC’s coupling with salience (insula) and reward regions (striatum) links self-relevant safety processing with affective and reward systems, suggesting its role in mediating the personal significance of stimuli and maintaining a sense of self across contexts [48]. This is also consistent with our prior work demonstrating that safety conferred through protection is distinct from threat despite both occurring in aversive contexts [49], likely due to its self-referential signaling. The PAG did not emerge as part of the safety network, consistent with its role in fast innate defensive reactions [13,50]. It did however emerge preceding risk of shock both during outcomes for lost battles and under increasing danger during Meta-representation.

Battle outcomes were tested as Safety Recognition and served as the purest test of safety neural circuitry. Neural activation in response to safety certainty, when subjects won the battle and were 100% safe from shock compared to when they lost battles, increased in the vmPFC, striatum, and hippocampus. In response to lost battles when there was risk of electric shock, compared with won battles, subjects demonstrated increased engagement of canonical defensive circuitry in the PAG and bilateral insula. Activation of the striatum and hippocampus indicates subjects learned during outcomes but did not reengage these circuits during prediction. The vmPFC, however, was engaged in response to both Safety Prediction and Recognition. We interpret the difference in vmPFC engagement compared with striatum and hippocampus engagement to indicate a more general role of the vmPFC in recognizing and predicting safety, rather than tracking outcomes to reinforce learning. This interpretation is further supported by using stimuli that had known and biologically relevant danger values (predators and weapons), which likely attenuates online associative learning during the task because the danger of the stimulus is already coded as part of its inherent properties [51]. The brain should treat these stimuli as known [52] and, therefore, vmPFC activation is unlikely a result of generating new learning effects in response to incidental shock pairings.

We identified a safety coding network that included subcortical and cortical regions involved in diverse processes including learning, reward valuation, and affect signaling. Although this study makes a significant advance in understanding how the brain contributes to estimating safety, many questions remain. First, how universal is the role of the vmPFC in coding safety? This study used a model of predator–prey interactions. Although humans are not typically exposed to predation, everyday threats induce neurobiological and psychophysiological states like those observed under predatory threats [3,13,53]. Using outwardly dangerous animals as threats eliminated interference from prior social experiences and allowed us to test the neural systems involved in Safety Value Updating. However, future work should examine safety coding during a diversity of threats including complex human interactions. Second, how do safety network communications evolve as a function of spatiotemporal dynamics? Our prior work shows that neural systems involved in defensive responding are dissociable along the threat imminence continuum [50,54]. The task used in the current study was not designed to examine dynamic threat nor did it evoke escape behavior. Further work is needed to determine under what conditions vmPFC-supported cognition is unavailable. Third, how do affect dynamics influence safety estimation? We did not collect participant report on their affective response to stimuli during the task. As a result, we cannot be certain whether our findings reflect an affective or non-emotional component of safety processing. Other work indicates a role of the vmPFC in the extinction of emotional arousal [55], and future work should consider how that role generalizes to safety estimation. Lastly, our MRI sample size prevented the examination of brain-based individual differences in age and psychopathology. Adolescence is a critical time to study safety computations given the prevalence of anxiety disorders, changes in metacognitive abilities, poorer threat-safety discrimination compared with adults, and imbalance in amygdala–vmPFC contributions to safety processing [24,56–59]. These features of adolescent development may result in impaired self-relevant safety processing.

The vmPFC coded safety during all task states from Safety Prediction to Meta-representation to Safety Recognition and Value Updating. Activation to these states was primarily identified in area 14m, with Safety Recognition and Value Updating extending to area 10. We identified what appeared to be a gradient from the posterior to the anterior part of vmPFC area 14m with activation extending more anterior as safety increased in certainty. This apparent gradient supports the possibility of subparcellation of area 14m into smaller areas with posterior and anterior distinction. During detailed mapping of vmPFC architecture, Mackey and Petrides note that “the posterior part of area 14m is less well developed and is less granular than the anterior part” [34]. This aligns with our recent proposition [2] that the posterior vmPFC may encode simpler representations of threat, whereas the anterior vmPFC may encode more complex safety associations. This apparent gradient aligns with the broader organizational gradient of the prefrontal cortex, where anterior regions are generally more granular and complex, reflecting their role in higher-order cognitive and integrative functions. Our findings support assertions that the anterior vmPFC plays a role in integrating the value of self-relevant stimuli to influence the higher-order construction of affective processes, including safety [2,60].

Beyond identifying how the human brain codes safety, our findings have potential implications for improving clinical interventions for anxiety. Current therapies focus on threat extinction but are ineffective for up to 50% of individuals [61]. A major problem with studying safety through the lens of threat extinction is the assumption that safety is the inverse of threat (in the absence of the aversive event the stimulus itself becomes “safe”). This confounding association does not consider how safety fluctuates independent of threat, for example, when protective resources can change safety while external threats remain unchanged. Current therapies target extinguishing fear responses to threats [62], but our data suggest focusing on self-relevant safety cues may be a promising therapeutic avenue. Also supporting a departure from extinction-focused approaches, recent work showed repetitive transcranial magnetic stimulation (rTMS) modulation of the anterior mPFC inhibited implicit fear reactions to learned threats [63]. This is a departure from emphasis on the dorsolateral PFC as a regulatory hub, which may be limited to extinction paradigms [64,65]. Intriguingly, the mPFC is a hub of the brain’s default mode network (DMN), which point to safety as an aspect of baseline human cognition. Psychopathologies, like anxiety, are often characterized by DMN dysfunction [66,67], which may mechanistically explain co-occurring deficits in safety estimation. Our findings provide a neuroscientifically grounded framework of safety beyond threat extinction and set the stage for future research to better understand how the human brain adaptively codes safety.

Finally, our work is situated in a broader understanding that the vmPFC supports self-referential processing, extending this role to safety estimation. The brain is engaged in considerable focus on tracking associations related to the self [31,32,35,46,48,68]. Our study demonstrates this does not just occur in conspecific social interactions or in identifying which objects belong to oneself, but also with respect to evolutionarily conserved threat appraisal. When the self is under threat, the vmPFC may help integrate threat-related stimuli into the broader narrative of self-representation, shaping responses based on past experiences and perceived safety. This raises a compelling case for considering how altered vmPFC activity in affective clinical disorders reflect disruptions in assessing and assigning personal significance to events, potentially tied to deficits in views of the self’s capacity to establish safety.

Hypotheses and methods were preregistered on the Open Science Framework (OSF), https://osf.io/hw3r9.